ARTIFICIAL INTELLIGENCE

AND CRIMINAL JUSTICE

December 2020
 SCIENTIFIC SUPERVISOR

ANTONELLA MASSARO
Associate Professor of Criminal Law
“Roma Tre” University

AUTHORS OF THE RESEARCH

ANGELO GIRALDI
Ph.D. student in Criminal Law
Murcia University - “Roma Tre” University

LORENZA GROSSI
Ph.D. student in Criminal Law
“Roma Tre” University

ANTONELLA MASSARO
Associate Professor of Criminal Law
“Roma Tre” University

LAURA NOTARO
Ph.D. student in Criminal Law
Pisa University

PIETRO SORBELLO
Field Officer in the “Guardia di Finanza” [Italian Financial Police]
Ph.D. in Criminal Law
 TABLE OF CONTENTS
Preface 7

INTRODUCTION
Antonella Massaro

1. Artificial intelligence and the criminal system: the reasons 9

for research
2. Contents of the research 9

CHAPTER I
ARTIFICIAL INTELLIGENCE, LAW AND CRIMINAL JUSTICE:
RECONSTRUCTIVE COORDINATES
Antonella Massaro

1. From the metaphor of law as a machine to the reality

of automated law 13
2. The possible relations between artificial intelligence and 16
criminal law
2.1 a) the delinquent algorithm 20
2.2 b) the investigating and/or fact-finding 22
algorithm
2.3 c) the consulting and/or judging algorithm 24
3. Courses, recourses and deviations in the parabola of 30
calculable law

3
 CHAPTER II
ARTIFICIAL INTELLIGENCE AND PREDICTIVE POLICING IN
THE MODERNISED INVESTIGATION STAGE
Angelo Giraldi

1. The usefulness of artificial intelligence in the

investigation stage 37
2. Predictive policing techniques: in the search of the
investigative “Paretian social optimum” 42
3. Algorithmic systems and availability of big data 45
4. Innovative crime prevention strategies: predictive
policing methods 49
4.1 Methods for predicting crime 52
4.2 Methods for predicting the future aggressor 53
4.3 Methods for drawing up a criminal identikit 55
4.4 Methods for predicting the future victim 57
5. Predictive policing and scientific analysis models 58
5.1 Models for identifying criminal hot spots 59
5.2 Spatiotemporal analysis and geospatial risk
factors 68
6. The experience of predictive policing in Anglo-
Saxon countries 69
7. The experience of predictive policing in Europe 73
8. Criticisms, risks and legal limits: towards a “shared”
standardisation of predictive policing 78

CHAPTER III
ARTIFICIAL INTELLIGENCE AND CRIMINAL JUSTICE
Laura Notaro

1. Experiences of application 86
1.1 The United States experience 88
1.1.1. Decisions on pretrial release 90

4
 1.1.2. Sentencing 92
1.2 European tendencies and experiences 94
2. Possible fields of application for AI in criminal
justice: a systematic overview 100
2.1 Assessment of the probability of repeat offending 100
2.2 Prediction of the decision-making result 106
2.3 Further theoretical fields of application 108
3. Criticisms and the risks of using AI technologies in the
criminal system 109
3.1 Risk-assessment tools 109
3.1.1. Some observations 117
3.2 Tools for predicting the decision-making result 119
3.3 Other fields of application 121
4. Legal limits on the entry of AI into the criminal system 121
4.1 The Ethical Charter of the CEPEJ of the Council of
Europe 123
4.2 European Union law: the limits to “decisions based
solely on automated processing” 127
4.3 The limits of domestic law: the ban on
criminological expert opinions 130
5. Final remarks. Spaces of application and precautions
for AI in criminal matters 135

CHAPTER IV
PREDICTIVE SOFTWARE AND CRIMINAL LAW
Lorenza Grossi

1. Introduction 141
2. What is artificial intelligence? 142
3. A “system” for criminal law 145
3.1 Predictive policing: Keycrime 147
3.2 Predictive justice: COMPAS 155
4. Areas of criticism and future prospects 161

5
 CHAPTER V
DATABASES AND PREDICTABILITY.
THE FINANCIAL RELATIONS ARCHIVE AND THE RISK
ANALYSIS FOR EVASION
Pietro Sorbello

1. Foreword 168
2. Police activity, the right to the protection of personal data
and the general interest of preventing and pursuing crimes
(hints) 170
3. The thread between economic analysis of law and
criminal policy, new technologies and criminal law
(and procedure) 177
4. The evolution of the regulation of the financial
relations archive. Analysis of the risk of evasion and
the interoperability of databases 180
5. The acquisition of notices of “theoretical” crime.
Final reflections 189

Bibliography 195

6
 Preface

Among the strategic technologies of the 21st century, so-

called artificial intelligence (AI), i.e. the use of intelligent algorithms
in the most varied sectors, is assuming an ever prominent role. This
concerns technologies that have undoubted potential and are
contributing to increasing efficiency and precision in the health,
agricultural and production sectors. However, they also imply
undeniable risks and potential critical issues, especially when you
want to apply them in particularly delicate fields which more
immediately involve individual rights and freedoms.
In more recent years, European institutions have paid growing
attention to the use and the implications of artificial intelligence: in
2016, the European Parliament examined, with different resolutions
on big data and robotics, the effects on, for example, education,
industrial production and the welfare of individuals. More recently,
the Commission presented a communication entitled “Building Trust
in Human Centric Artificial Intelligence” (8th April 2019) and
published a white paper entitled “Artificial Intelligence - A European
approach to excellence and trust” (19th February 2020).
In October 2020, the European Parliament adopted an
important “package” of resolutions and reports regarding civil
responsibility on artificial intelligence, intellectual property rights for
the development of artificial intelligence technologies, the ethical
aspects of artificial intelligence, robotics and related technologies. The
common theme running through these texts is the need for the use of
these tools and technologies to always be in full compliance of the
fundamental rights and that the principles of transparency and non-
discrimination are always ensured.
One area where this need is most pressing is undoubtedly law
and justice, and criminal law and justice in particular. A draft
resolution, currently under scrutiny by the Committee on Civil
Liberties, Justice and Home Affairs, of which I am Member, is
dedicated to the use of artificial intelligence by police and judicial
authorities in the criminal domain.

7
 The profile of the relations between artificial intelligence and
criminal law and justice is particularly complex, because of the
multiple modes in which the algorithms can come into prominence:
like tools involved in various ways in the commission of crimes, or as
tools supporting the activities of the public minister or the forces of
law and order (so-called predictive policing) or still as a tool
supporting the judicial body (so-called predictive justice).
This clearly concerns a sector of the system in which
individual guarantees and the principles of the state of law must be
upheld with the greatest vigour. To this end, a careful analysis that
goes on from understanding how these algorithms work and probes the
potential benefits, but also the potential problems to test, is therefore
necessary, always without prejudice to the individual’s right not to be
subjected to decision-making based solely on automated processing,
which may be capable of producing an adverse legal effect on the data
subject or in any case significantly affect him or her (Article 11 of
Directive 2016/680).
This is the aim of this study which, taking into consideration
the predictive policing and justice tools currently used in both member
states of the European Union and third party countries, highlights the
potential and limits of using artificial intelligence in the criminal
justice field, in an attempt to define the spaces for application and
identify the necessary precautions.
I would therefore like to thank Prof. Antonella Massaro and
the Department of Jurisprudence at the “Roma Tre” University for
their contribution offered to the debate on such a strictly topical issue.

Caterina Chinnici

8
 INTRODUCTION
Antonella Massaro

Table of Contents: 1. Artificial intelligence and criminal justice: the reasons for
research. – 2. Contents of the research.

1. Artificial intelligence and criminal justice: the reasons for

research

The “Artificial Intelligence and Criminal Justice” research,

carried out on commission of the Hon. Caterina Chinnici of the
European Parliamentary Group of Socialists and Democrats, considers
the dense and diverse web of relations founded on the use of
“intelligent algorithms” and the criminal system.
The initial scepticism that the possible use of artificial
intelligence (also) arouses on the criminal law front has been replaced
by the gradual and inevitable acknowledgement of how the use of tools
based on ever more refined and accurate algorithms is now a tangible
reality and is being progressively consolidated.
The aim, therefore, is to highlight the critical issues that
artificial intelligence undeniably entails, without nevertheless giving
in to a priori and apodictic approaches that deny any virtuous potential
for the tools in question.

2. Contents of the research

The research has also tried to “untie” the artificial intelligence-

criminal law pair in order to better specify the possible applications of
algorithms in the “criminal” law sector.

9
 In particular, the attention was focused on the two poles of
predictive policing and predictive justice. The first includes the
systems that are put forward to predict the committing of a crime
and/or identify the author of an already known series of crimes through
the statistical processing of data, whilst predictive justice refers to
software designed as a tool to aid the judge, used in both pretrial
release and sentencing.
The contribution of Angelo Giraldi, in particular, examines
the predictive policing models. The term predictive is broken down
into its dual meaning of prediction and investigation: this concerns two
strictly connected dimensions, both based on the “discovery” of
similarities and analogies found in the analysis of variables that are
constantly related to each other.
The universe of predictive justice was, on the other hand,
examined by Laura Notaro, who moving on from an analysis of the
main software used in the US and European experience, examines the
prognostic sentences that could be affected by the systems in question:
not only the sentences regarding repeat offending, but also, for
example, those on which the granting of probation is based, the
application of security measures (based on a social danger decision)
or, even, those which are dotted with the discipline of penal execution.
The software that allows (or would allow) the judicial outcome to
become more predictable is also examined, again examining their
limits and potential.
Lorenza Grossi has performed a more detailed examination of
some software which, though for different reasons, can be considered
paradigms of the predictive policing and predictive justice models.
The choice, in particular, concerned Keycrime and COMPAS.
Keycrime is the predictive policing system used, since 2008, over the
Municipality of Milan area and, since 2009, over the whole Milan
district, whilst COMPAS is the predictive justice software developed

10
in 1998 and used in the United States of America. COMPAS
(Correctional Offender Management Profiling for Alternative
Sanctions) was developed in 1998 by the private company
Northpointhe Inc. (now Equivant). The purpose of the analysis is not
only to show the strong and weak points of systems that reproduce the
operation of those examined, but also to highlight the different
“criticality rate” that predictive policing systems entail as compared to
predictive justice ones.
Finally, the contribution of Pietro Sorbello examines the
modes by which the use of increasingly complex and interconnected
databases can support or, in certain cases, replace the traditional search
for notices of crime: analysis of metadata, i.e. of the relationships
between individually insignificant data, is in effect a particularly
useful tool in certain sectors which, like that of tax crimes, lend
themselves more clearly to the use of predictive tools.
The research has highlighted how the use of artificial
intelligence is a response perceived as “opportune”, especially by
those who emphasise the dysfunctions of the current criminal system.
The demand for security, the phenomenal complexity taken on by
many spheres of crime, the biases that the human being runs into and
the elephantiasis of criminal proceedings are all factors that speak
volumes for an increasingly marked “objectification” of the rules of
the game, to make the response of the criminal system more efficient,
but also more controllable.
To act as a counterpart to this reassuring vision, there is, first
of all, the need for greater transparency of the reference algorithms,
which, designed by private entities without a “public sharing” of the
source code, risk entering into collision with the most elementary
requirements on which the state of law is built. To this it can be added
that at least some predictive policing software is based on a causal
determinism that: human actions, on certain conditions arising, are

11
predictable, pace free arbitration and the criminal law of
responsibility. Attention is shifted from the deed of the author,
suggesting a change of paradigm (at least partial) compared to the one
offered by classical free matrix criminal law. A calculable law is
perhaps finally marked as a welcoming priority, especially in a
historical context that appears dominated and absorbed by a law with
an unstable and perennially changing case law source: nevertheless,
the risk to be highlighted (and avoided) is the one of a system that may
be reduced to the tired and stagnant repetition of itself, losing that
capacity to adapt “spontaneously to the changed safeguard
requirements, according to perhaps less intelligent mechanisms, but
certainly more vital.

12
 CHAPTER I

ARTIFICIAL INTELLIGENCE AND CRIMINAL JUSTICE:

RECONSTRUCTIVE COORDINATES

Antonella Massaro

Table of Contents: 1. From the metaphor of law as a machine to the reality of

automated law. – 2. The possible relations between artificial intelligence and criminal
law. – 2.1. a) the delinquent algorithm. – 2.2. b) the investigating and/or fact-finding
algorithm. – 2.3 c) the consulting and/or judging algorithm. – 3. Courses, recourses
and deviations in the parabola of calculable law.

1. From the metaphor of law as a machine to the reality

of automated law

The metaphor of law as a machine belongs to centuries of

western culture and politics 1, being in fact used in different contexts
and for different purposes. If, one hand, comparing law to a machine
means highlighting the profile of its artificiality, the metaphor in
question, on the other hand, is tied to the dual thread of the
Enlightenment ideals of calculable law, based on a sure legislation and
a judge resized to the role of “mouthpiece of the law”. The vision of
the law as a computer, capable of operating in binary code
(lawful/unlawful, permitted/prohibited) and structured according to
criteria of formal rationality, also comes close to this vision2.

1
G. ITZCOVICH, Il diritto come macchina. Razionalizzazione del diritto e
forma giuridica in Max Weber, in Materiali per una storia della cultura giuridica,
31/2, 2001, 379, also for the necessary references.
2
The three “meanings” of the comparison between law and machine to which
reference is made in the text are those highlighted by F. POGGI, Il diritto meccanico.

13
 Nevertheless, the “law and machine” pair is anything but an
undisputed image: indeed, no hesitation has been made in maintaining
that “the metaphor of the law as a machine is among the most
unfortunate that have ever been used”3.
Even Max Weber, when he observed that modern capitalism
would need law that could be computed like you do with a machine,
acknowledged the widespread fear amongst the operators of the law
that they would be reduced to mere automatons, who would have seen
the removal of both their moral independence and the social dignity of
the social class they belonged to: the progressive spread of realist, anti-
dogmatic trends or, in any case, ones that propose an alternative to
legal positivism, would also arise from the requirement of judges to
free themselves from the image of “automatic dispenser of the law”
(Subsuntionsautomat)4, a mere acknowledger of positive law in the
context of that process of bureaucratisation of case law which, once
again, is functional to a law that can really be considered calculable5.
The objections to a “mechanical law” model go beyond the
questions regarding the concrete practicability of an interpretation that
is merely (or mainly) declaratory and pertain, substantially, to the risk
of a not-so-desired deresponsibilisation of the judger-automaton: the
application of the law would require at every step management

La metafora del diritto come macchina e i suoi limiti, in Dir. quest. pubbl., 9/2009,
395-396.
3
G. TUZET, Sul possibile moto della macchina, in Dir. quest. pubbl., 9/2009,
359. For the debate aroused by the work of Tuzet, see G. ITZCOVIC Sulla metafora del
diritto come macchina, cit., 379 et seq..; M. NARVÁEZ MORA, Metaforear, in Dir.
quest. pubbl., 9/2009, 385; F. POGGI, Il diritto meccanico, cit., 395 et seq..; the
response of G. TUZET, Il diritto non è una macchina, in Dir. quest. pubbl., 401 et seq.
4
M. WEBER, Economy and society, IV, Edizioni di Comunità, 1995, 79.
5
With regard to Weber’s idea of the machine as a “repressed spirit” and,
therefore, an already fully unfurled rationalisation process, with the consequent
“resizing” of the role of the judge, G. ITZCOVICH, Il diritto come macchina, cit., 372-
373 (spec. note 30) and 391 et seq.

14
implemented by persons capable of self-control and therefore
responsible6 or, if you will, judges (not inanimate but) with a soul
capable of bringing “the great weight of this enormous responsibility
that is doing justice”7. Without counting that flesh and blood men are
fallible beings, but also capable of learning from their mistakes or, in
any case, “modernising” interpretive solutions or argumentative
models that are no longer fit for the growing complexity of the deeds
that the law is called upon to regulate8.
Precisely at a historical time when, also in civil law systems,
we are seeing an increasingly marked case law component in criminal
law (we should not hold back in speaking about a “true case law
source”, in certain cases), the idea of mechanical law is returning to
monopolise the theoretical and political debate. Nevertheless, it no
longer concerns passionate discussions around a more or less effective
and/or shareable metaphor, but prospects of investigation aimed at
probing the practicality of justice governed literally by machines or,
better, with the aid of artificial intelligence and algorithms that are
becoming progressively refined.

6
G. TUZET, Sul possibile moto della macchina, cit., 359.
7
P. CALAMANDREI, Processo e democrazia, in Opere giuridiche, I, Roma TrE-
Press, 2019, 650.
8
See M. WEBER, Economy and society, III, Edizioni di Comunità, 1980, 41
and 170, where it is observed how one can object to being “outside life” under law
based on merely logical/rational constructions, given that the solutions indicated by it
are often placed in an irrational relationship with the operators of the economic traffic
who are called upon to operate in situations with a growing complexity of conflicts of
interest.

15
 2. The possible relations between artificial intelligence
and criminal law

The generic reference to artificial intelligence tools that could

take the place of judges is in fact shown to be broadly insufficient,
given that phenomena which are very different to each other and not
always easily understood by the layman are hidden behind the label of
artificial intelligence.
The concept of artificial intelligence, at first approximation,
refers to those machines that are capable of thinking or rather imitating
human thought 9. John McCarthy, who is attributed with being the
father of the expression artificial intelligence, based his studies on the
premise that every manifestation of learning may be described so
precisely as to allow a machine to simulate it: the problem of artificial
intelligence would thus essentially consist of obtaining a machine that
behaves in ways that would be defined as intelligent if a human being
were to behave in such a way10. It is also specified that artificial

9
Obviously, the background to this definition echoes the known distinction
between strong artificial intelligence, according to which machines have a mind and
are therefore intelligent, and weak artificial intelligence, according to which artificial
intelligence is a mere imitation, not a duplication of human intelligence. This concerns
a distinction which, as is known, goes back to J.R. SEARLE, Minds, brains and
programs, in Behavioral and Brain Sciences, 1980, 417: “According to weak AI, the
main value of the computer in the study of the brain lies in the fact that it provides us
with a very powerful tool: it allows us, for example, to formulate and verify
hypotheses in a more precise and rigorous way. On the other hand, according to strong
AI, the computer is not simply a tool for studying the mind, but rather, when suitably
programmed, is a real mind; i.e. it is possible to affirm that computers, once given the
right programs, literally understand and possess other cognitive states. For strong AI,
since the programmed computer possesses cognitive states, the programs are not
simple tools that allow us to verify psychological explanations: the programs are
themselves those explanations”. Amplius J. KAPLAN, Intelligenza artificiale. Guida al
futuro prossimo, Luiss University Press, 2017, 74 et seq.
10
J. MC CARTHY, M.L. MINSKY, N. ROCHESTER, C.E. SHANNON, A Proposal
for the Dartmouth Summer Research Project on Artificial Intelligence, 1955, available
at www-formal.stanford.edu.

16
intelligence consists in the capacity of the machine to make
appropriate generalisations promptly and on a limited data base11, and,
it could be added, in such way that the human mind is capable of
recognising the rationality and, consequently, the reliability of the
process implemented.
Without undervaluing the ethical and philosophical
implications of the question Can machines think? which Alan Turing
asked himself in 1950 in the context of his Imitation Game12 and
continues to resonate in the investigations on artificial intelligence 13,
even legal type ones, it is clearly a question that can be left unsolved.
By way of provocation, but not too much, the answer could be that
asking yourself if machines can think is as relevant as asking yourself
if submarines can swim14 or, for that matter, that aerospace
engineering certainly does not put itself forward as creating “machines
that fly like a pigeon in such a perfect way as to deceive the flight of a
pigeon”15. Despite the provocations, the concept of “thought” could
perhaps be replaced by that of “reasoning”, which is more “objective”
and less prone to misunderstanding. Machines are capable of
reasoning and they are also capable of doing it in the forms of legal
reasoning.
We are dealing with affirmations that appear even clearer after
the development, in the broadest framework of artificial intelligence,
of “machine learning” and “deep learning” systems. The reference is

11
J. KAPLAN, Intelligenza artificiale, cit., 2017, 19.
12
A.M. TURING, Computing machinery and intelligence, in Mind, 1950, 433.
13
See, for example, O. DI GIOVINE, Il Judge-bot e le sequenze giuridiche in
materia penale (intelligenza artificiale e stabilizzazione del precedente), in the
process of being published, § 3.2.
14
E.W. DIJKSTRA, The Threats to Computing Science, 1984, available at
www.cs.utexas.edu.
15
S. RUSSEL, P. NORVIG, Artificial Intelligence. A modern approach, I, II ed.,
Pearson 2005, 6.

17
to those machines that are capable of “learning from experience” and,
therefore, improving themselves autonomously: one of the possible
forms of machine learning is represented by deep learning, where the
learning is automatic via the use of artificial neural networks, delving
right to the bottom of the pathway for comparing the human brain and
the mechanical one. Among the more familiar applications of deep
learning are translation, voice recognition and computer vision 16.
The daily familiarity and, therefore, the “normality” that each
person can experiment with regard to the most recent developments in
artificial intelligence is a constant datum in dissertations on the
subject, even those called upon to investigate the potential and limits
of a possible interaction between artificial intelligence and law 17.
Similarly, the cinematographic citations are particularly frequent.
Here artificial intelligence, to stay within the dynamics starting from
the “cops and robbers” dichotomy, is called upon to play the role of
the baddie (think of Terminator)18 or the goody (like Robocop)19, even
punishing the guilty before they have even committed a crime (like the
Precrime system in Minority Report20). There is no shortage of refined
literary citations, ranging from the pages of a “general” author like Ian

16
For a brief but effective overview, N. BOLDRINI, Deep Learning, cos’è
l’apprendimento profondo, come funziona e quali sono i casi di applicazione, in
www.ai4business.it, 9th August 2019.
17
J. NIEVA-FENOLL, Intelligenza artificiale e processo, translated by P.
Comoglio, Giappichelli, 2019, 9; F. BASILE, Intelligenza artificiale e diritto penale:
quattro possibili percorsi di indagine, in Dir. pen. uomo, 29th September 2019, 2;
18
D. BEN-HARI, Y. FRISCH, A. LAZOVKI, U. ELDAN, D. GREENBAUM, Artificial
Intelligence in the Practice of Law: An Analysis and Proof of Concept Experiment, in
Richmond Journal of Law and Technology, 2/2017, 10.
19
F. BASILE, Intelligenza artificiale e diritto penale, cit., 9.
20
C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale: molte
speranze e qualche equivoco, in Dir. pen. cont., 6/2019, 48.

18
McEwan (Adam. Machines like me)21 to those of a lawyer-fiction
writer like Pierre Janot (Lex Humanoïde, des robots et des juges)22.
Even the European Parliament Resolution of 2017 does not refrain
from placing itself in that line of ideal that goes “from the monster
created by Mary Shelley to the classical myth of Pygmalion, passing
though the story of Golem of Prague and the robot of Karel Čapek” 23.
Without renouncing the taste for cinematographic citation, it
could be said that the legal debate is questioning itself on the times and
methods (the year seems to less in discussion) in which the jurist alter
ego of HAL (the intelligent “heuristic algorithm that pilots the
Discovery 1 spaceship on 2001-A Space Odyssey) will govern the fates
of a criminal justice, the weight of which is becoming increasingly
burdensome for fragile human shoulders. Even if we indulge the taste
for a citation taken more from the popular imagination of eighties TV
series than the cinema, one could ask oneself when the times will be
ripe for opening the doors of courts and/or lawyers’ offices to Small
Wonder stories, the intelligent robot which “imitates” human
behaviour, which reads millions of pages in a few seconds and stores
the content and which should not feel emotions but ends up
establishing a very close relationship with its “little human brother”.
The unstoppable technological evolution-revolution that has
made science of what was science fiction, which was really a scenario
reserved, until several decades ago, for the artistic creativity of

21
O. DI GIOVINE, Il Judge-bot e le sequenze giuridiche in materia penale, §
3.1.
22
E. RULLI, Giustizia predittiva, intelligenza artificiale e modelli
probabilistici. Chi ha paura degli algoritmi?, in Analisi giuridica dell’economia,
2/2018, 535.
23
Resolution of the European Parliament of 16th February 2017, on
recommendations to the European Commission on Civil Law Rules on Robotics
(2015/2103(INL)) Introduction A.

19
“visionary” intellects, exercises an undeniable fascination, especially
on the generation of “analogue natives”.
Nevertheless, on closer inspection, it certainly serves no
purpose to recall the tender look of Haley Joel Osment in Steven
Spielberg’s AI to exorcise the uncontrollable apocalyptic scenarios
arising from an invasion of robot-jurists. Much more prosaically, it
concerns making a point on the current uses of artificial intelligence
to tackle and/or resolve legal points in order to outline a situation that
is reasonably expected in the near future, by evaluating its costs and
benefits. It mainly concerns checking whether a “more calculable” law
than the current one can arise from the concrete realisation of a law
that functions (not like a machine, but) through a machine.
The possible intersections between criminal law and artificial
intelligence belong to very different sectors, even if these converge in
a broader system viewpoint.
For purely descriptive-classification purposes, one can
distinguish between the delinquent algorithm, the fact-finding
algorithm and the consulting and/or judging algorithm.

2.1. a) the delinquent algorithm

The expression “delinquent algorithm” summarises the

question regarding Artificial Intelligence-Supported Crimes and,
therefore, artificial intelligence as a tool involved, in a different
manner, in committing certain crimes: think of the different types of
“computer” crimes, financial fraud and drug trafficking but also
drones guided to kill and guarantee urban safety or, even, accidents
caused by cars with an automatic pilot24.

24
R. BORSARI, Intelligenza Artificiale e responsabilità penale: prime
considerazioni, in www.medialaws.eu, 3/2019, § 2; T. KING, M. AGGARWAL, M.

20
 The prospect for investigation on this front is (at least)
twofold.
In first place, the increasingly widespread use of artificial
intelligence “for criminal purposes” as well lets new methods of crime
emerge which the judge and legislator are called upon to deal with, to
verify whether the traditional cases of crime are enough to meet the
“new” protection requirements or whether, conversely, interventions
capable of taking the potential offence of the criminal algorithm into
due consideration have become necessary.
Secondly, it would seem that the traditional paradigm of
criminal responsibility, based on the crime committed by the natural
person, is preparing itself to undergo the umpteenth aftershock: after
the discussions aimed at overcoming the dogma of societas delinquere
non potest, it concerns up to what point the principle of machina
delinquere non potest25 can be valid, given that, in certain cases, the
model of “indirect responsibility of the person” would seem not to be
adaptable to the more complex matters, especially those that see
machine learning mechanisms involved 26.
Again, in the context of the “persons of the crime”, it could
even be conjectured that the algorithm is involved, not as active
subject, but as potential victim. The machine-person equalisation,
especially on the feelings and emotions front, should lead one to
consider robots as possible holders of legal interests to be protected:

TADDEO, L. FLORIDI, Artificial Intelligence Crime: An Interdisciplinary Analysis of

Foreseeable Threats and Solutions, in Science and Engineering Ethics, 2019, 1 et
seq.; F. BASILE, Intelligenza artificiale e diritto penale, cit., 24 et seq.
25
A. CAPPELLINI, Machina delinquere non potest? Brevi appunti su
intelligenza artificiale e responsabilità penale, in Discrimen, 27th March 2019, 1 et
seq., which also underlines this through the fallacy of the parallel with the
responsibility of legal entities.
26
R. BORSARI, Intelligenza Artificiale e responsabilità penale, cit., § 4. See
also, F. BASILE, Intelligenza artificiale e diritto penale, cit., 29 et seq., to which you
are referred for wide bibliographical references.

21
if, for the criminal algorithm, the most immediate parallelism is the
one with the criminal responsibility of entities, for the victim
algorithm, one can often recall the possible analogy with the protection
of animals, which have been progressively raised from the reductive
qualification of “objects” to “sentient beings” and, thus, possible
“persons in law”27. Can you imagine the criminal significance of
sexual violence committed against an android? Or the abuse (not
simple damage) of a robot that gives assistance to disabled persons 28?

2.2. b) the investigating and/or fact-finding algorithm

The uses of artificial intelligence as a tool in support of the activities

of the public minister and the criminal police refer firstly to the so-
called predictive policing software, which, by re-processing an
amount of data that a human operator would be incapable of managing,
guides in the identification of hot spots, meaning places where the risk
of a crime being committed is high, or gives support in crime linking
activities, predicting the future commission of crimes by determined
persons or ones that have yet to be determined29.
To this one can add programs that are of aid in reconstructing
facts starting from clues: STEVIE, for example, constructs stories that
are consistent with the data provided to it, whilst ALIBI assesses the

27
For the implications that the process in question entails with regard to the
criminal protection of animals, A. MASSARO, I reati “contro gli animali” tra
aspirazioni zoocentriche e ineliminabili residui antropocentrici, in Cult. dir., 1-
2/2018, 79 et seq.
28
In problematic form, F. BASILE, Intelligenza artificiale e diritto penale, cit.,
32-33.
29
All the necessary references in F. BASILE, Intelligenza artificiale e diritto
penale, cit., 10 et seq.;

22
alternative explanations for the behaviour of a person charged with a
crime30.
Among the “home grown” applications, the example offered
by KEYCRIME is particularly well known. This is a program
developed by the policeman Mario Venturi and currently in use in the
Milan police headquarters. The data show that the use of KEYCRIME
has not only caused a significant reduction of certain crimes (mainly
robberies), but may also have supplied significant aid in the solving of
crimes that have already been committed31. The peculiarity of the
software lies in the fact that it is also capable of processing data that
pertain to the behavioural profile of the author, increasing, for
example, its level of accuracy compared to similar algorithms when it
is dealing with the future commission of crimes by a certain person.
The possible risks are fairly clear. First of all, the police could
act on the basis of biased machines, meaning like machines
programmed on the basis of prejudices, especially ethnic and racial
ones, which leads to intensifying the controls and repression in certain
areas in particular and with reference to certain categories of persons 32:
a machine would certainly be quicker and more efficient that a human
operator, but not, for example, less racist that those who program it or
(above all) those who are called upon to “manage” its indications. This
is without taking into account the questions regarding the processing

30
Amplius E. NISSAN, Digital Technologies and Artificial Intelligence’s
Present and Foreseeable Impact on Lawyering, Judging, Policing and Law
Enforcement, in AI&Soc., 2017, 450 et seq.
31
R. PELLICCIA, Polizia predittiva: il futuro della prevenzione criminale?, in
www.cyberlaws.it, 9th May 2019.
32
An effective overview of the question is offered by R.M. O’DONNEL,
Challenging Racist Predictive Policing Algorithms under the Equal Protection
Clause, in New York University Law Review, 2019, 544 et seq.

23
of personal data and also the invasiveness of the police in public
spaces33.

2.3 c) the consulting and/or judging algorithm

Further “legal” applications of artificial intelligence are those

that pertain to a possible use of the algorithm in a consulting and/or
judging function, making it an instrument capable of not only being
put to the service of the judge, but also of conferring a more suitable
implementation of the right to defence.
With regard to the consulting algorithm, there are at least two
implementations that can come into consideration. These are tied by a
relationship of (at least partial) mutual implication:
a) first of all, artificial intelligence allows you to
organise ever more sophisticated databases, which do not
represent a mere deposit of data and information, but are also
capable of guiding the user in the search for relevant
regulatory and case law references or, perish the thought,
formulating “opinions”.
That traditional databases have revolutionised the way
operators in the law work is so obvious a fact that there is no need for
further considerations. The “most evolved” tools to which reference
has been made are increasingly more common and are being refined.
ROSS INTELLIGENCE, the legal version of IBM’s more well-known
Watson Debater and already christened “SIRI for the law”, is for
example capable of legal advice starting from an illustration of fact
expounded in “common English”: it suggests case law references and

33
J. KREMER, The End of Freedom in Public Places? Privacy problems
arising from surveillance of the European public space, University of Helsinki, 2017,
available at helda.helsinki.fi.

24
quotations of doctrine and, since it “learns” from passed interactions,
every response of ROSS increases its precision for the future. And the
list could go on34.
b) the consulting algorithms, especially those
that exploit machine learning or even deep learning, thus
represent the first step towards the objective of what is
increasingly labelled as “predictive justice”35: i.e. the
algorithm would allow you to predict, with an ever increasing
level of reliability, the outcome of a trial.
Particularly well-known, also for its “positive” results, is the
algorithm developed by a research group, which has shown itself
capable of “calculating” the outcomes of the sentences of the European
Court of Human Rights with an accuracy level of 79%36. The
researchers specify that, whilst previous studies gave value to
metadata capable of influencing the votes of the judges (gravity of the
crime, political positions of the judges and indications from amici
curiae), their study proposes to predict the decision of the case basing
itself solely on information of a text nature37. Moreover, the result of
the study seemingly shows how the contribution of the “fact” as
opposed to the “law” is determining, in accordance, according to the
authors, with the indications coming from legal realism 38.

34
For all the necessary references, D. BEN-HARI, Y. FRISCH, A. LAZOVKI, U.
ELDAN, D. GREENBAUM, Artificial Intelligence in the Practice of Law, cit., 31 et seq.
35
E. RULLI, Giustizia predittiva, intelligenza artificiale e modelli
probabilistici, cit., 539, deems the use of the adjective “predictive” inappropriate,
since it risks evoking an unreal “divination” regarding the outcomes of a trial.
36
N. ALETRAS, D. TSARAPATSANIS, D. PREOTIUC-PIETRO, AND V. LAMPOS,
Predicting judicial decisions of the European Court of Human Rights: a Natural
Language Processing perspective, in PeerJ Computer Science, 2016, 1 et seq.
37
N. ALETRAS, D. TSARAPATSANIS, D. PREOTIUC-PIETRO, AND V. LAMPOS,
Predicting judicial decisions of the European Court of Human Rights, cit., 2.
38
N. ALETRAS, D. TSARAPATSANIS, D. PREOTIUC-PIETRO, AND V. LAMPOS,
Predicting judicial decisions of the European Court of Human Rights, cit., 11 et seq.

25
 Leaving aside the implications of a “theoretical” nature that
could be drawn from the study in question, what is important to
underline is that the high degree of accuracy refers to the breach/non-
breach pair of a certain ECHR article: a prediction, in other words, that
concerns the decision-making device not its explanations.
Finally, we reach the judging algorithm. This probably
concerns the application of artificial intelligence that makes the
greatest “impression” on the collective imagination and not just on
those working the field.
Artificial intelligence can actually put itself at the service of
the judge in a considerable number of cases, not fully attributable to
the idea of the robot/sentence mouthpiece.
For example, amongst the algorithms put at the service of the
judge, those which are capable of calculating the risk of the crime
being repeated and thus can be used as a basis for applying
precautionary measures or alternative measures to imprisonment and
as a tool to refer to when commuting the sentence.
It is precisely the latter application that has made the feats of
COMPAS (Correctional Offender Management Profiling for
Alternative Sanctions) known outside the United States and this has
become the emblem of the advantages, but mainly the limits, of
artificial intelligence in a judging function. COMPAS applies
algorithms created on the basis of numerous personal data of the
offender, but the private company that drew up and marketed the
system has not fully revealed its mechanism and operation. In the
celebrated case of Loomis, the defendant disputed the use of COMPAS
at sentencing, both because of the lack of transparency that featured in
its operation and because the algorithm was programmed on the basis
of racial and gender prejudices. The Supreme Court of Wisconsin
highlighted the possible risks inherent in the use of COMPAS, which
essentially consisted of the lack of transparency of its operating

26
mechanisms, the application of logics processed on a collective basis
(referring to groups or classes of people) and not strictly individual
and the possible overestimate of the risk of some ethnic minorities
committing crimes: notwithstanding this, the Court deemed lawful a
proportioning of the sentence based also, but not solely, on the
indications supplied by COMPAS39. The question, in matters like this,
regards not so much the use of the algorithm, but its programming or,
more exactly, the recruitment and checking of the technicians asked to
process and manage the algorithm itself.
It is not by chance that both the transparency, taken as
accessibility, comprehensibility and external verifiability of the
computerised processes used in the court and the prohibition on
creating or aggravating, through algorithms, discrimination against
groups or individuals are fundamental principles identified in the
“European ethical charter for the use of artificial intelligence in
judicial systems and their environment”, drawn up, in the context of
the Council of Europe, by the European Commission for the
Efficiency of Justice (CEPEJ)40.
Artificial intelligence could also intervene when evaluating
evidence41, with appreciable prospects not only in documentary

39
Wisconsin S.C., State v. Loomis, 881, N.W.2d 749 (2016). Amplius J.
NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 56 et seq.; F. BASILE,
Intelligenza artificiale e diritto penale, cit., 19 et seq.; S. CARRER, Se l’amicus curiae
è un algoritmo: il chiacchierato caso Loomis alla Corte Suprema del Wisconsin, in
Giurisprudenza Penale Web, 4/2019.
40
The other principles are those regarding fundamental rights, the quality and
security of data, the checking by the user of the results produced by a machine: for an
effective analysis of the European Charter, see S. QUATTROCOLO, Intelligenza
artificiale e giustizia: nella cornice della Carta etica europea, gli spunti per
un’urgente discussione tra scienze penali e informatiche, in Leg. pen., 18th December
2018, 1 et seq.
41
Fundamental on the point, L. L. LUPÀRIA, Prova giudiziaria e ragionamento
artificiale: alcune possibili chiavi di lettura, in Il concetto di prova alla luce

27
evidence but also in reference to statements made by the parties or
witnesses: this would involve programming the algorithm so that it
could valorise, for example, the more convincing acquisitions on the
psychology of evidence or, perhaps, make use of neuroscience to
distinguish untruthful statements from truthful ones 42.
The algorithm could then be of aid to the judge to test the
reasonable exclusion of alternative reconstructions to the one that
would lead to affirmation of the defendant’s responsibility, even if the
artificial intelligence applied to the evidential standards shows a
structural deficiency, since it concerns tools more directed to the future
than to the reconstruction of events in the past 43.
On top of the questions regarding evidence, in the broad sense,
the use of the judging algorithm that arouses major interest in the
perspective of calculable law is undoubtedly the one of possible
application in the true activity of sentencing. The base idea is of a
machine which, far from representing a mere aid to the judge, almost
fully replaces him/her in drafting the sentence, at least in the so-called
easy cases.
The main objection that could be raised against a robot fully
drafting a sentence would seem to lie in the structurally persuasive
nature of legal reasoning, which, as the radical criticisms of the
syllogism of an Enlightenment die, would not be reduced to mechanics
and the always identical application of a criminal case in point, on the

dell’intelligenza artificiale, edited by J. Sallantin, J.J. Szczeciniarz, Giuffrè, 2005,

XIV et seq.;
42
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit.,70 et seq. V. anche
A. TRAVERSI, Intelligenza artificiale applicata alla giustizia: ci sarà un giudice
robot?, in Quest. giust., 10th April 2019, § 6.
43
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 95 et seq.

28
contrary requiring evaluations which, at least currently, a machine is
not capable of performing44.
If, however, we are dealing with easy cases 45, or rather very
easy ones, such as for which, once the pre-trial activity is carried out,
application of the regulation to the concrete case is found to be “taken
for granted”, it could be imagined that the sentence is automated, but
at this point the most natural outcome would be the disappearance of
the reasoning. It has been recently maintained that if the decision can
be taken by an algorithm, the function of which is known and the
accuracy of which can be trusted, then the reasoning would become
superfluous 46.
The algorithm, for that matter, is not only capable of reasoning
like a human being, but can be programmed to exclude the heuristics
and cognitive biases which, almost inevitably, condition human
behaviour, even that of a judge 47.
In conclusion, the algorithm judge would be quicker, more
efficient, with less prejudices, capable of taking predictable decisions
and, therefore, bringing a significant contribution (at least) to a

44
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 105 et seq. On
computer tools capable of providing an argumentative support, see E. NISSAN, Digital
Technologies and Artificial Intelligence’s Present and Foreseeable Impact, cit., 448-
449. For a critical analysis of the application of current artificial intelligence tools to
true “judging” activities, see C. TATA, The Application of Judicial Intelligence and
‘Rules’ to Systems Supporting Discretionary Judicial Decision-Making, in Artificial
Intelligence and Law, 1998, 203 et seq.
45
The (possibility of a) distinction between easy cases and difficult cases to
delimit the possible areas of application for artificial intelligence in the court is the
premise that moves O. DI GIOVINE, Il Judge-bot e le sequenze giuridiche in materia
penale, cit., § 0.
46
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 108 e 134 for the
implications on accusations.
47
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 32 et seq.; O. DI
GIOVINE, Il Judge-bot e le sequenze giuridiche in materia penale, cit., § 2.

29
stabilisation of case law. Briefly: who is afraid of artificial
intelligence?

3. Courses, recourses and deviations in the parabola of

calculable law

On first sight, it would seem that we are in front of an

umpteenth “course and recourse” of history which does not have the
characteristics of an authentic “revolution”, but rather that of a return
to the origins. Artificial intelligence, at least in some of its possible
applications, would seem to re-propose the “old and outdated” idea of
legal syllogism, confirming that certain ways of understanding the
human reasoning applied to law are so rooted as to let some hints
emerge again from the “precious ingenuity” of Beccaria48.
Between the idea of the judge-automaton as a metaphor of
Enlightenment die and automated-law as a real application of new
generation artificial intelligence systems, not only would clear points
of contact seem recognisable, but also just as many macroscopic
aspects of divergence.
The point of contact is represented by the declared reaction to
a case law chaos that is superficially perceived as intolerable.
The age of codification represented a turning point or, at least
a new start, synthesised not so much by the introduction of something
new into the system, but in the ordered rationalisation of the previous
legal tradition. In the age of codification, the code reservation recently
introduced in Article 3-bis of the Criminal Code sounds more like the

48
The reference is to the title of V. VELLUZZI, La preziosa ingenuità:
Beccaria, lo spirito della legge e il sillogismo giudiziale, in Riv. int. fil. dir., 2014, 687
et seq.

30
desperate impotent attempt to steer the future legislator, though
through the weak weapon of legislative self-constraint, without any
possibility of remedying the damage, in part inevitable, caused by the
past legislator49. The cries of Manzoni have left the place to clashes
which, in the space of a few weeks, are recorded in the same section
of the Court of Cassation, with the united sections called to ever more
frequent and closer interventions that are ever less stable.
The solution that the legal Enlightenment proposed, in order
to amend some of the more conspicuous “defects of case law”, was to
marginalise the judge, marking his/her difference to the legislator and
confining him/her in the role of innocuous bureaucrat. The primary
interest of the enlightenment followers was not to systematise the law
or, even less, to understand the inexplicable way that jurists reason:
they more simply wanted to abolish it, replacing it with the law and
bureaucrats called to apply it50.
The possible uses of artificial intelligence in a criminal law
context (but not only) would seem to go from an apparently mirror-
like premise: the aim would not be to marginalise the judge and his/her
work, but to offer him/her a precious instrument of assistance which,
in the final analysis, means exalting the role. The “European Ethical
Charter on the Use of Artificial Intelligence in Justice” also
emphasises the necessary control by the user over the results supplied
by the machine, given that the objective of computational tools applied

49
For a detailed critical scrutiny of the code reservation introduced with
Article 3-bis of the Criminal Code, see La riserva di codice (art. 3-bis c.p.) tra
democrazia normante e principi costituzionali. Apertura di un dibattito, in Leg. pen.,
20th November 2018, 1 et seq.
50
M. BARBERIS, Cosa resta del sillogismo giudiziale? Riflessioni a partire da
Beccaria, in Materiali per una storia della cultura giuridica, 1/2015, 163.

31
to criminal justice should be that of enhancing, not reducing, the
decision-making autonomy of the user51.
The “profession of judge” is currently suffocated and
annihilated by an elephantine set of bodies and procedures, which are
placed as fearful adversaries to case law stabilisation52. If judges could
dedicate themselves to “serious matters”, delegating the fate of
chicken and/or snack thieves to fast, economical and implacable
machines, the stabilisation of case law would become a progressively
more defined objective53.
Flesh and blood judges, for that matter, act mechanically in
the majority of cases: the more the workload becomes onerous, the
more the decisions become standardised, the more recourse to the
“short cuts” of heuristics increases 54. Writing innumerable routine
documents slows down the system and increases the risk of error55.
If then you cannot manage to deflate the criminal process
though the traditional paths (decriminalisation and process reform), it
is worth so much intervening upstream the phenomenon, launching a
lifeboat which never tires, does not pretend to do “doctrine” with its
opinions and (perhaps even because of this) makes less mistakes, to
the judge who is drowning in a sea of processes.

51
S. QUATTROCOLO, Intelligenza artificiale e giustizia, cit., 9 underlines the
centrality of these premises, also, in the context of the Ethical Charter; the concept of
user is used ambiguously, referring at times to justice professionals and at other times
to the intended recipient of the decision.
52
O. DI GIOVINE, Il Judge-bot e le sequenze giuridiche in materia penale
(intelligenza artificiale e stabilizzazione del precedente), cit., § 0
53
The alternative amongst serious things and less serious things, which was
provocatively referred to in the text, would actually be translated into one between
easy cases and difficult cases, as for example suggested by O. DI GIOVINE, Il Judge-
bot e le sequenze giuridiche in materia penale, cit., § 1.
54
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 32-33.
55
K. BRATING, J.C. LESTER, C.B. CALLAWAY, Automating Judicial Document
Drafting: A Discourse-Based Approach, in Artificial Intelligence and Law, 1998, 114.

32
 It is not the task here to evaluate the effective practicality of
the uses of artificial intelligence, to which a nod has been made, but it
is sufficient to pause on what would seem to be the highest price to
pay in terms of system. The risk, almost unanimously underlined, is
one of a stagnation of case law56, hermeneutic plastering57, with a
consequent heterogenesis of purposes: artificial intelligence should
contribute to the evolution of the system, not hobble it with “intelligent
fetters”.
For that matter, close to those who exalt the role of the judge
aided by a machine, one cannot fail to observe that judges and their
courts will be increasingly unnecessary58 and that lawyers, for their
part, will become an endangered race59.
On closer inspection, the most interesting prospects on the
relations between judge and machine are not those that do not limit
themselves to predicting man being accompanied by an algorithm or
those that insist on the increasingly broader prospects of a “self-
sufficient” judge-robot, but those according to which the effect of
artificial intelligence applied to the criminal process can influence the
selection of the operators in the law: not only the judges, but also the
lawyers.
The use of artificial intelligence will perhaps make a smaller
number of judges necessary, but what is certain is that there will have
to be highly specialised professionals capable of understanding and

56
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 21.
57
C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale, cit., 51.
58
D. BEN-HARI, Y. FRISCH, A. LAZOVKI, U. ELDAN, D. GREENBAUM, Artificial
Intelligence in the Practice of Law, cit., 35.
59
D. BEN-HARI, Y. FRISCH, A. LAZOVKI, U. ELDAN, D. GREENBAUM, Artificial
Intelligence in the Practice of Law, cit., 37. For the impact of artificial intelligence on
the practice of the legal profession, see also P. MORO, Intelligenza artificiale e
professioni legali. La questione del metodo, in Journal of Ethics and Legal
Technologies, 2019, 24 et seq.

33
connecting the results supplied by the machine60 and capable of
bringing real added value in solving those “difficult cases” that cannot
be fully delegated to the machine. Not to mention, it is worth stating,
that, at least as a rule, the label of “easy case” or “difficult case” can
only be attributed ex post, making a preventive filter, that is not always
so foregone and automatic, necessary.
The best antidote to a possible stiffening of the system, then,
is the presence, at the time of impugnment and above all in the Court
of Cassation, of judges with the capacity not to be satisfied always and
regardless with the result supplied by the machine: “the main criterion
for the selection of the members of last resort bodies must be their
creativity”61.
Similar considerations, mutatis mutandis, are valid for
lawyers. The lawyer will have to know how the machine operates, not
to mention that predictive algorithms, as stated, also represent a valid
instrument for strengthening the right to defence. The lawyer must
therefore be capable of drawing up strategies that can “resist the
algorithm” when, for example, the circumstances of the case impose it
or when the machine has to settle into the tired repetition of
anachronistic solutions and/or ones that are systematically
unreasonable.
We are dealing with a scenario which, even before being
projected into the future, should lead to opening our eyes on the
present, starting from the current mechanisms for the recruitment of
law professionals. It is written, for example, that the universities
should keep pace with artificial intelligence and that future jurists must
be taught how to relate with algorithms right from the start.
Nevertheless, if it is really necessary to increase the value of the
creativity of judges and lawyers to make the difference, it is paramount

60
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 21.
61
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 135.

34
that the university lecture rooms should contribute to allowing that
creativity to emerge62.
It could be speculated that the obsolescence of machines will
shortly leave space for that of human beings and that the radical
transformations already in progress in the world of work cross the
border and have a final landing where “there is no use for people” 63.
Nevertheless, given that people are still of use to the state, the
presumption of modelling men on machines (after the apparent
presumption of modelling machines on men) would be hasty, feeding
a spiral that would make man ever more replaceable: the jurist, for
his/her part, would have to be trained first and then selected giving
value to precisely those capacities and attitudes that allow him/her to
use the machine as a tool and be differentiated from an algorithm,
however evolved it may be.
The open frontiers of the unrelenting vertiginous development
of intelligent algorithms seem to dust off the dream of every jurist,
philosopher or thinker that pursued the pipe-dream of a calculable law:
a machine is capable of calculating law, confirming the fact the
calculability is not burnt out in either the superstructures which
bourgeois ideology uses to “justify” capitalism or in the naive
conviction of those who take refuge in ancient legal mythologies to
prevent case law from taking its final revenge over the law.
Nevertheless, the stake is very high. The case law change, if it
does not cross the border into the unmanageable chaos of a case law

62
See D. BEN-HARI, Y. FRISCH, A. LAZOVKI, U. ELDAN, D. GREENBAUM ,
Artificial Intelligence in the Practice of Law, cit., 39: “Law schools will change
dramatically, not least because we will need fewer lawyers. Moreover, the nature of
legal learning will change to include subjects that are not taught in law schools today-
creativity, understanding of statistics, big data analysis, and more”.
63
The reference is clearly to the title of J. KAPLAN, Le persone non servono.
Lavoro e ricchezza nell’epoca dell’intelligenza artificiale, translated by Ilaria
Veronica Tomasello, Luis University Press, 2016. See also, K. SCHWAB, La quarta
rivoluzione industriale, Franco Angeli, 2016.

35
that struggles to find its (albeit temporary) stabilisation, represents a
completely physical phenomenon which, on the contrary, ensures an
“automatic” adjustment of the legal system to the social, cultural and
technological changes. If, for the sake of argument, the result of a
“calculable law” in the strict sense were reached, the quid pro quo
would be a substantial stagnation of the law: the case law would repeat
itself almost identically to itself, with the risk of leading to the
devaluation of the peculiarities of the single cases and reaching
standardised formulae for managing legal points.
We are clearly dealing with a risk that is inversely proportional
to the role which is intended to be reserved for the “human mind” as
opposed to the activity carried out by the algorithm and which,
therefore, tends to be attenuated in those models that imagine the
support of artificial intelligence in the sectors where its intervention
could contribute to reducing the margin of error and/or ensuring a
higher degree of objectivity in sentences that would otherwise be
delegated to an inscrutable subjectivity. On the other hand, different
conclusions could be reached if we were to imagine freeing some
sentences from the intervention of the “human judge”, even
speculating on the obsolescence of the obligation for reasoning and the
confrontational system.
To this we can add that, whatever the contribution of
intelligent algorithms that can be imagined with reference to the law
and criminal justice, the question of a necessary control of the
programming parameters must always be asked: as long as we are
dealing with tools managed “technically” by private entities, there
will, necessarily, be a problem of upholding some of the fundamental
principles on which the state of law is founded.

36
 CHAPTER II

ARTIFICIAL INTELLIGENCE AND PREDICTIVE POLICING

IN THE MODERNISED INVESTIGATION STAGE

Angelo Giraldi

Table of Contents: 1. The usefulness of artificial intelligence in the investigation stage.

– 2. Predictive policing techniques in the search of the investigative
“Paretian social optimum”. – 3. Algorithmic systems and availability of big
data. – 4. Innovative crime prevention strategies: predictive policing
methods. – 4.1. Methods for predicting crime. – 4.2. Methods for predicting
the future aggressor. – 4.3. Methods for drawing up a criminal identikit. –
4.4. Methods for predicting the future victim. – 5. Predictive policing and
scientific analysis models. – 5.1. Models for identifying criminal hot spots.
– 5.2. Spatiotemporal analysis and geospatial risk factors. – 6. The
experience of predictive policing in Anglo-Saxon countries. – 7. The
experience of predictive policing in Europe. – 8. Criticisms, risks and legal
limits: towards a “shared” standardisation of predictive policing

1. The usefulness of artificial intelligence in the

investigation stage

One of the most useful and productive uses of artificial

intelligence undoubtedly regards the first stage of the criminal process,
which is made concrete in the investigation activities. The
investigative stage, in fact, is the fulcrum of criminal justice: on one
hand, the activity of predictive investigation is functional for the
security of the individual and the collective, in its intent to effectively
identify planned crimes or, in any case, future ones; on the other hand,
the operations of the authorities tasked with the post factum

37
investigation are functional to tracing the conduct back to its author
and possibly the criminal organisation behind him/her.
The scientific and technological progress that is sweeping
across today’s society is also the harbinger of innovation in the
predictive and investigative sector, which characterises the stage in the
criminal proceedings under analysis64. Under this perspective, the idea
that decisions, both pretrial and in court, should be based on the
maximum rigour possible is gaining more ground.
From this particular point of view, an albeit brief
acknowledgement of the heuristic methodologies dominating the
thought of the judicial operator can be useful: this person, grappling
with the different forms of crime, must carry out the investigation, be
he/she a police officer or a public minister, in the exercise of his/her
function.
In the search for the facts and the truth, as well as one
regarding the various sources of investigation, the tasked person uses
a methodology that is undoubtedly seeped with subjective
characteristics which, almost unanimously, are defined as risky.
Since the dawn of social and philosophical thought, the myth
of justice remains anchored to criteria of impartiality and objectivity
which, in practice, encounter grey areas where the objectivity is
pervaded by outside influences. One cannot omit to say that the same
juristic heuristics, as social science, is practised (up to now) by
physical persons, whose reasoning is also irrationally determined by

64
For an in-depth examination of the theme of artificial intelligence as
ultimate purpose of criminal law, see C. BURCHARD, L’intelligenza artificiale come
fine del diritto penale? Sulla trasformazione algoritmica della società, in Riv. it. dir.
proc. pen., 4/2019, 1908 et seq.: “Intelligent algorithms are capable, through
technology, of tackling daily problems, to be benefit of everybody, that nevertheless
go beyond human capabilities; and certainly better, more quickly and more
economically than human decision-makers”. See also V. MANES, L’oracolo
algoritmico e la giustizia penale: al bivio tra tecnologia e tecnocrazia, in Intelligenza
artificiale. Il diritto, i diritti, l’etica, edited by U. Ruffolo, Giuffrè, 2020, 547 et seq.

38
personal psychological factors that have nothing to do with the fact of
the crime committed or to be committed.
An interesting study in this regard dates back several years 65.
Eight Israeli judges were examined over a period of ten months, for a
total number of fifty daily sittings. The individual judge, in those
occasions, was asked to decide on requests for conditional release put
forward by convicts in different prisons. The days were divided into
three moments: firstly, the early hours of the morning, in which the
judge started his/her work; secondly, the second part of the morning,
which started after a refreshment break; and finally the rest of the
working day after lunch.
The intention was to relate the judicial decisions with the times
that they were adopted, in an attempt to understand how much those
decisions could be attributed not just to the discretionary interpretation
of the magistrate – always anchored to the canons of law – but, rather,
to his/her psychological situation, completely independent of the legal
cases.
The result that the scholars reached was of obvious importance
as it emerged, clearly and eloquently, that the percentage of petitions
for conditional release granted at the start of every time frame in the
day was by far greater than those dealt with once half of the time span
in question had passed. In fact, if the favour towards conditional

65
The study is briefly reported in the introduction to a contribution on the
reasoning of judges made by science logicians and philosophers. The original study is
however to be attributed to “Danziger, Levav and Avnaim-Pesso (2011, see figure 1
at p. 6890 in particular), and then taken up by many others, including Kahneman
(2011, pp. 48-9), who saw to its publication on PNAS (the prestigious revue of the
American National Academy of Science). The most common interpretation, also
suggested by the authors, is that the progressive deterioration of the judges’ cognitive
resources determines a tendency to choose the default option, i.e. in favour of the
status quo. Significant critical comments on the details of study, especially on its
interpretation, are in Weinshall-Margel and Shapard (2011) and Glöckner (2016)”.
See G. CEVOLANI – V. CRUPI, Come ragionano i giudici: razionalità, euristiche e
illusioni cognitive, in Criminalia, 2017, 181 et seq.

39
release was around 65% of the petitions dealt with in the early hours
of the morning, it tended move to zero as time went by, until a rest that
was consistent with the refreshment break that the magistrate took.
Once the next time frame started, there were no changes and the
percentages, initially high, of granting the benefit dropped drastically
as time went by.
Essentially, it clearly emerged from the study that tiredness,
fatigue, the need for sustenance and the tedium of judicial authority
were risk factors. In fact, the convicts’ chances varied considerably
according to psychological/physical status of the decision-maker.
The above-mentioned study, although emblematic, is only a
part of the numerous studies on human reasoning, which, obviously,
can also be applied to a court sitting. Reasoning and human logic are
analysed by a growing number of scholars, the data for which are not
in the least comforting. The mental characteristics of the individual
persons reveal that the basic mechanisms of our cognitive and
decision-making activities (also with regard to the most banal daily
activities) constantly deviate from what is prescribed, abstractly, by
the so-called theories of correct reasoning.
This cannot necessarily be attributed, like in the study as of
above, to mixed factors of human nature. Circumstances completely
extraneous to the person, which cannot be predicted by him/her, can
also influence his/her decision in one way or another.
This is the case, for example, with anchoring, which was the
subject of an experiment on a group of German jurists with judicial
experience behind them66. The former magistrates were proposed a
case of theft, the fruit of the authors’ imagination, on which they have
to make a judgement. The persons participating in the study were

66
B. ENGLICH – T. MUSSWEILER – F. STRACK, Playing Dice With Criminal
Sentences: The Influence of Irrelevant Anchors on Experts’ Judicial Decision Making,
in Personality and Social Psychology Bulletin, 32/2006, 194.

40
divided into two groups, each of which received an accurate
description of the case together with two loaded (unbeknown to them)
dice.
Each of the groups were asked to throw the dice to get, by
adding up the results, the number corresponding to the months of
conviction requested by the prosecution. In the first group, the throw
of the dice was programmed so that a result of three months was
obtained (so-called low anchor), corresponding to a relatively
restrained request for conviction by the prosecution. In the second
group, however, the figure obtained was greater and added up to nine
months (so-called high anchor).
The result of the study, though significant and satisfying from
a scientific point of view, was rather disappointing from a legal point
of view. The first group, which was supplied with a low anchor,
completely unpredictable from the participants’ point of view, opted
to convict the defendant to five months of imprisonment. The second
group, on the other hand, having received a high anchor, set the
penalty at eight months. As shown by the analysis carried out, the
anchoring effect played a fundamental role in the decision-making
process of the persons involved. The fact that all the people in the first
group had calibrated the sentence for an amount considerably lower
than the participants in the second group could not be attributed to a
purely causal mechanism.
It would seem that the external influence on the decision-
makers may be attributed to the reliability which the prosecution often
generates on the judge, because the result of the throwing of the dice
was considered as a proposal for conviction by the Public Prosecutor.
Yet, the participants – people with an education level above the
average, and already invested with jurisdictional powers – were fully
aware that, the prosecution was represented by an inanimate object
totally incapable of thinking and providing rational proposals. In this

41
sense, that the participating jurists trusted (so-called anchoring) an
external, inanimate and completely random element is astounding: the
judgement of the professionals in question was broadly conditioned by
numbers that were apparently alien to matters.
In the wake of these exemplary experiences, artificial
intelligence intervenes to mitigate the subjective influences of the
physical persons who operate in the world of justice, in an attempt to
remove, as objectively as possible, the biases that pervade criminal
proceedings67.

2. Predictive policing techniques: in the search of the

investigative “Paretian social optimum”

The idea that decisions, because they potentially refer to every

individual, must be based on rigorous analysis of data through
scientifically recognised methods has already taken hold in the current
doctrine68. Thus, data processing, appropriately collected and
analysed, is reaching almost uncontrollable quantitative levels. At the
base of the artificial intelligence programs, including predictive and
investigative software, lie big data, characterised in the first place by
the huge quantity69. It is not by chance that the algorithms on which

67
On the problem of prejudices (biases), see amplius O DI GIOVINE, Il judge-
bot e le sequenze giuridiche in materia penale (intelligenza artificiale e stabilizzazione
giurisprudenziale), in Cass. pen., 3/2020, 951 et seq.
68
The theme of data analysis and its use in scientific terms is very topical in
the contemporary and international doctrinal debate. See ex plurimis, N. HART – M.
YOHANNES, Evidence Works: Cases Where Evidence Meaningfully Informed Policy,
in Bipartisan Policy Center, 2019, available in full at the following address:
https://bipartisanpolicy.org/report/evidenceworks/.
69
For an in-depth examination of the concept of big data, see A. BONFANTI ,
Big data e polizia predittiva: riflessioni in tema di protezione del diritto alla privacy
e dei dati personali, in Rivista di diritto dei media, 3/2018.

42
said computerised applications of mathematics and statistics are more
efficient the more data they possess.
The final aim of the algorithm is to know an initially unknown
value, to be reached through the analysis and systematisation of a large
set of individual data70. In order to better understand the purpose of
using this (originally) mathematical method, we must ask ourselves
about the range of the “unknown value” that we are looking for
through examination of the data available to the operators.
Firstly, the value referring to an event that has already
happened can be defined as unknown, because there is (and it could
not be otherwise) always an element that the interpreter is not capable
of knowing with sufficient certainty. On first sight, one could exclude
from the category of past events those for which an audiovisual
recording is available, capable of providing the interpreter with the
possibility, ex post, of entering virtually into the spatiotemporal
context of the fact. Yet, even in this case, there would be elements that
cannot be known, since human action, though faithfully reproduced
with digital tools ad hoc, hides a series of indeterminate psychological
processes which currently elude every type of recording or
understanding by third parties that can be said to be absolutely beyond
doubt. Secondly, and for obvious reasons, every type of present and, a
fortiori, future event falls into the category of unknown values.
On the basis of the observations made, it is appropriate to
analyse the term “predictive” in more detail. It is often used as an
attribute referring to software and policing algorithms, meant in the
broad sense as increasingly effective tools for predicting and
preventing crime.
In fact, if one of the aims of artificial intelligence is to find
“unknown values” referring to future and chance (hence the term

70
F. PROVOST – T. FAWCETT, Data Science for Business: What you need to
know about data mining and data-analytic thinking, O’Reilly Media, 2013.

43
predictive) events, one cannot overlook the relationship between the
new technological tools and facts belonging to the past and present. In
this perspective, predictive policing algorithms - and the related
software - must not solely mean the methods exploited to estimate the
probability of future crimes, but rather also those adopted to aggregate
the useful data to bring clarity on crimes that have already been
committed and on the continued behaviours perpetrated in the
present71.
Thus, the term predictive must be considered in its double
meaning of prediction and investigation. In the first case, predictive
policing will be aimed at anticipating future crimes; and, in the second,
it will act as the main character in the search for the author of the crime
or, in broader terms, the attribution of the latter to a superordinate
organisation. It should be noted that the two proposed meanings are
strictly interconnected: the investigative activity and the
characteristics of the facts and authors flow into the data used for
predicting future crimes; vice versa, the successful outcome of the
strictly predictive activity greatly facilitates subsequent investigation
operations.
The true usefulness of so-called predictive policing lies in the
discovery of similarities and analogies following the analysis of
variables that are constantly related to each other. In this sense, the
change of pace is notable compared to the traditional, in a manner of
speaking, tools tied to “natural” intelligence, which have been applied
right up to contemporary times 72. The margin for reducing crime is
expanded with the new heuristic methodologies tied to the science of

71
The Organization for Security and Co-operation in Europe bears witness to
this in its report Guía de la OSCE sobre actividad policial basada en la inteligencia,
available in several languages at the address
https://www.osce.org/es/chairmanship/455536.
72
See amplius, G. TUZET, L’algoritmo come pastore del giudice? Diritto,
tecnologie, prova scientifica, in MediaLaws, 2019.

44
probability and the development of technology. This is mainly due to
the fact that the resources, though limited, and the data held by the
tasked authorities are exploited in an optimised manner as compared
to the past. In fact, on one hand, a similar use of resources allows,
mutatis mutandis, an investigative “Paretian social optimum” to be
achieved; on the other hand, it allows the criminal police, in the broad
sense, to organise more efficient operating strategies and decision-
making.
After all, it is not by chance that important studies in the field
of environmental psychology have shown an albeit regularity of
delinquency; in fact, it seems that criminals tend to act in known
places, without travelling afar and in determined or determinable times
or places73.

3. Algorithmic systems and availability of big data

An indispensable condition for drawing up effective strategies

and reliable predictions is the availability of data74. This condition is,

73
On this point, see S. VEZZADINI, Profilo geografico e crime mapping. Il
contributo della criminologia ambientale allo studio del delitto, in Scena del crimine
e profili investigativi. Quale tutela per le vittime?, edited by R. Bisi, Franco Angeli,
2006; P.J. BRANTINGHAM – P.L. BRANTINGHAM, Environmental criminology: from
theory to urban planning practice, in Studies on Crime and Crime Prevention, 7/1998;
J. ECK – S. CHAINEY – J. CAMERON – M. LEITNER – R. WILSON, Mapping Crime:
Understanding Hot Spots, in National Institute of Justice, 2005, available at the
following address: https://nij.ojp.gov/library/publications/mapping-crime-
understanding-hot-spots.
74
Data is the fulcrum on which artificial intelligence turns. Obviously, the
data must firstly be captured using so-called sensors. They “could be cameras,
microphones, a keyboard, an Internet site or other data entry systems, as well as
sensors of physical quantities (e.g. temperature, pressure, distance or force/torque
sensors or tactile sensors)”. The data collected are often “digital data and there is an
immense availability; and with regard to the data, it must immediately be stressed that
the quality of the final result depends, to a large extent, on logical correctness and the
completeness of the data collected; on the other hand, if the data used to feed or train

45
at the same time, one of the most significant limits of artificial
intelligence that exists today. The big data held by the authorities are
nothing but resources produced by the authority itself: a vicious circle
is thus set up, given the (possible) partiality of the information and
data found by those to whom such data will return in an artificially
aggregated and analysed form.
Nevertheless, although this allows considerable criticism to be
advanced against artificial intelligence, it is difficult to imagine an
artifice that operates otherwise. The only (im)practicable way would
be to aggregate both sets of data coming from, firstly, the empirical
comparison carried out in the local reality and, then, from the results
of an ideal “perfect” society. It is difficult to even imagine a complete
study that describes, in terms of absolute precision, the perfect social
system; even if it was theorised, it would perhaps be unreliable.
In any case, given the possibility of solely using data regarding
real society, a fruitful collaboration between the different
professionals operating in social sciences becomes necessary (and
favourable)75. The world of justice is of course the first to be involved:
the investigations of the criminal police, just like the processes already
taught, are collectors of an inestimable amount of data that can be

an AI system are distorted, in the sense that they are not sufficiently balanced or
inclusive, the system will not be capable of generalising correctly and could adopt
unjust decisions that could favour some groups over others”. With the baggage of
resources thus identified, all that remains is to make two consequential steps:
“reasoning or the processing of information[, which] is a process operated through an
algorithm that acquires said data as input to then propose an action to be taken in light
of the object to be achieved”; at the end, the aggregation of data. To do this “the AI
system carries out the pre-chosen action through the actuators at its disposal, which
can be both software and physical elements (e.g. articulated arms and automatic
wheels); the latter capable of intervening, modifying it, on the surrounding
environment”. Thus F. BASILE, Intelligenza artificiale e diritto penale: quattro
possibili percorsi d’indagine, in Diritto Penale e Uomo, 29th September 2019, 6.
75
This hope is shared by J.H. RATCLIFFE, Intelligence-Led Policing,
Routledge, 2016.

46
validly used. Nevertheless, other important parameter sectors, which
can bring a volume of qualitatively and quantitatively different data,
must not be undervalued.
Think, for example, of the world of health in its broad sense,
which also includes the multi-form area of the social services in it.
These are the sectors which, also thanks to the implementation of
information technology, deal daily with information regarding
criminal or socially dangerous persons. In some ways, the health world
data can be a quid pluris compared to the information collected by the
criminal police, for different sets of reasons. Firstly, at a quantitative
level, entry into a further branch of social activities inevitably allows
the exponential increase of data. Secondly, the type of information
collected, e.g. from the social service, is by far different and certainly
complementary to what was found by the criminal police.
Given that there are data – or rather, big data – at the basis of
the whole artificial intelligence process, it is maybe opportune to
wonder about the very concept of “intelligence” which is often put to
the test by the methodology applied.
In this regard, the authoritative definition of artificial
intelligence provided by the independent group of high level experts
appointed by the European Commission for carrying out the
consultative function must be remembered76. For the EU experts,
“artificial intelligence (AI) systems are software (and possibly also
hardware77) systems designed by humans that, given a complex goal,

76
Information on the expert group appointed by the European Commission
can be found on the institutional website, at the following address:
https://ec.europa.eu/digital-single-market/en/high-level-expert-group-artificial-
intelligence.
77
The systems based on artificial intelligence can be considered both software
and hardware. More specifically, vocal assistants, programs for the analysis of images,
search engines and biometric recognition systems are well-known examples of
software. On the other hand, the implementation of such programs in peripherals that
act outside the digital world makes it that artificial intelligence is also present in

47
act in the physical or digital dimension by perceiving their
environment through data acquisition, interpreting the collected
structured or unstructured data, reasoning on the knowledge, or
processing the information, derived from this data and deciding the
best action(s) to take to achieve the given goal. AI systems can either
use symbolic rules or learn a numeric model, and they can also adapt
their behaviour by analysing how the environment is affected by their
previous actions. As a scientific discipline, AI includes several
approaches and techniques, such as machine learning (of which deep
learning and reinforcement learning are specific examples), machine
reasoning (which includes planning, scheduling, knowledge
representation and reasoning, search, and optimization), and robotics
(which includes control, perception, sensors and actuators, as well as
the integration of all other techniques into cyber-physical systems)” 78.

hardware. Think, for example, of driverless cars, drones robots and the various
applications of the Internet of Things. In this regard, see Communication from the
Commission to the European Parliament, the European Council, the Council, the
European Economic and Social Committee and the Committee of the Regions.
Artificial Intelligence for Europe, 2018, 2, available in the full version at the following
address: https://ec.europa.eu/digital-single-market/en/news/communication-artificial-
intelligence-europe.
78
Artificial intelligence (AI) means “Software systems (and possibly
hardware) designed by humans that, given a complex objective, act in the physical or
digital dimension by perceiving their environment through data acquisition,
interpreting the structured or unstructured data collected, reasoning on knowledge or
processing the information derived from this data and deciding the best actions to be
taken to achieve the given objective”. AI systems can use symbolic rules or learn a
numerical model and can adapt their behaviour by analysing how the environment is
influenced by their previous actions. As a scientific discipline, AI includes different
approaches and different techniques, like automatic learning (of which deep learning
and reinforcement learning are specific examples), mechanical reasoning (which
includes the planning, programming, representation of knowledge and reasoning,
research and optimisation) and robotics (which includes control, perception, sensors
and actuators, as well as the integration of all other techniques into the cyber-physical
systems”.

48
 Without prejudice to the definition proposed, it is appropriate
to make the following clarifications 79. Firstly, artificial intelligence,
because it was created by human beings, does not fully correspond to
the “naturalness” of the latter’s reasoning. Secondly, and not for
importance, the expression “artificial intelligence” is far from the
operation of the human mind, the “intelligence” of which remains to
this day an undetermined concept. It seems that “little, apart from
speculation and a naive way of thinking, connects today’s work in the
AI field to the mysterious mechanisms of the human mind; at least at
this stage, we are dealing with an engineering discipline with relations
with biological organisms that are more metaphorical and
“inspirational” than anything else80. For this reason, one often prefers
to replace the expression “intelligence” with “rationality”, where
rationality means the “capacity to choose the best action to take to
achieve a certain objective in light of some criteria for the optimisation
of the available resources”81.

4. Innovative crime prevention strategies: predictive

policing methods

If we examine the dynamics of predictive policing in more

detail, in the already-mentioned double meaning of prediction and
investigation, we can analyse the different strategies for preventing

79
The clarification that follows has been advanced by authoritative doctrine.
SEE BASILE, Intelligenza artificiale e diritto penale, cit., 5; C. TREVISI, La
regolamentazione in materia di Intelligenza artificiale, robot, automazione: a che
punto siamo, in Medialaws, 2018, 1 et seq.
80
J. KAPLAN, Intelligenza artificiale. Guida al futuro prossimo, Luiss
University Press, 2018, 81 et seq.
81
F. BASILE, Intelligenza artificiale e diritto penale, cit., 5.

49
crime and the functional tools for analysing the facts of an already
committed crime or one that is in progress.
Based on the models drawn up at a theoretical level, it would
be possible to predict when, where and how the crimes will be
committed, so that an attempt can be made to anticipate the causal
mechanism of the crime82. The ultimate aim of the predictive policing
method is a concatenation of elements that, together, make the
investigation stage more efficient. The increase in resources to
safeguard public safety in the areas at greatest risk allows us to identify
the areas where the criminal risk factors are concentrated and,
consequently, channel the intervention of the forces of law and order
to conduct targeted operations.
In the diagram in the figure that follows (see Figure 1), one
can distinguish the continuous interaction between the following
elements: data collection, analysis, police operations and criminal
response. It is complex to imagine the operational efficiency of
predictive policing without one of the mentioned components:
Data collection, suitably aggregated (fusion), is functional for
their analysis. The latter is prefixed with the objective of preventing
future crime, which inevitably guides the police operations. As a
consequence of these interventions, the criminals react in a diverse
manner, resulting overall in two fronts: on one side, the reactions of
the people in question solicit the requirement for new assessment and
consequently new operating actions; on the other, they create an
altered environment. In the latter case, the circle is closed because the
new information found on the so-called altered environment (i.e. the
data on the criminal (re)actions) return to be collected and aggregated.

82
See A. DI NICOLA – G. ESPA – S. BRESSAN – M.M. DICKSON – A.
NICOLAMARINO, Metodi statistici per la predizione della criminalità. Rassegna della
letteratura su predictive policing e moduli di data mining, in eCrime Working Papers,
2/2014, 5 et seq.

50
 The use of similar predictive models, therefore, allows police
forces to organise the resources they have so that they can repress,
thwart and, if possible, anticipate criminal conduct.

Figura 1 – Predictive policing method. Source: Predictive Policing.

The Role of Crime Forecasting in Law Enforcement Operations , in
RAND. Safety and Justice Program, 2013, XVIII.

The interventions carried out by the police forces achieve a

fairly high specificity and concreteness rate and finish up by being
remarkably targeted. The crime in certain areas could therefore be
considerably reduced, which would allow us to assess the operation of
artificial intelligence positively.

51
 The current state of the art seems to allow the classification of
the predictive policing strategies into four macro-categories83.

4.1. Methods for predicting crime

A first subset of predictive policing is represented by the

“methods for predicting crime” (see Table 184). All the types of
strategy aimed at predicting the place and time a crime will most likely
be committed can be attributed to it. It seems that this category has
been widely experimented with, particularly in the Anglo-Saxon
countries, where the final objective ends up by coinciding, in a certain
sense, with general prevention. In fact, as we will see, these operating
models, which benefit from the collection and processing of large
quantities of data, usually from crimes committed previously, are
prefixed with the purpose of anticipating (i.e. preventing) the criminal
event. On the other hand, different methods aspire to anticipating, and
even preventing, the repeat of it or a different crime by certain persons.

83
Ibid., 6.
84
“Table 1.1 summarizes predictive policing methods related to predicting
crimes, i.e., identifying places and times that correspond to an increased risk of crime.
As the table shows, conventional approaches start with mapping crime locations and
determining (using human judgment) where crimes are concentrated (the hot spots).
These approaches might include making bar graphs showing when crimes have
occurred (time of day or day of the week) to identify “hot times.” The corresponding
predictive analytics methods start, at the most basic level, with regression analyses
similar to what one would learn in an introductory statistics class and extend all the
way to cutting-edge mathematical models that are the subjects of active research.
Some methods also attempt to identify the factors driving crime risk”.

52
 Conventional Crime Analysis Predictive Analytics
Problem (low to moderate data (large data demand
demand and complexity) and high complexity)

Identify areas at increased risk

Advanced hot spot
Using historical crime Crime mapping (hot spot
identification models;
data identification)
risk terrain analysis
Using a range of Regression,
Basic regression models created
additional data (e.g., 911 classification, and
in a spreadsheet program
calls, economics) clustering models
Accounting for Assumption of increased risk in
increased risk from a areas immediately surrounding Near-repeat modeling
recent crime a recent crime
Graphing/mapping the
Determine when areas
frequency of crimes in a given Spatiotemporal analysis
will be most at risk of
area by time/date (or specific methods
crime
events)
Finding locations with the
Identify geographic features greatest frequency of crime
Risk terrain analysis
that increase the risk of crime incidents and drawing
inferences

Table 1 – Methods for predicting crime. Source: Predictive Policing. The Role of
Crime Forecasting in Law Enforcement Operations, in RAND. Safety and Justice
Program, 2013, 10.

4.2. Methods for predicting the future aggressor

The second category of the predictive methods consists of the

“methods for predicting the future aggressor” (see Table 285). Groups

85
“Table 1.2 summarizes methods to identify individuals at high risk of
offending in the future. The bulk of these methods relate to assessing individuals’ risk.

53
of individuals most exposed to the risk of delinquency are identified
through the strategies under analysis. However, in order to do this, it
is necessary to collect, analyse and aggregate the data for already
committed crimes, but the data concerning the individual authors of
the crimes must be added to that information. As can be deduced, one
of the main objectives of the methods falling under this category is the
reduction of repeat offending. Although theoretically there is not an
exact correspondence between committing a crime and its reiteration
(in the more or less distant future), in practice a high percentage of
specific reoffending has been found.

Conventional Crime
Predictive Analytics
Analysis (low to
Problem (large data demand
moderate data demand
and high complexity
and complexity)

Manual review of incoming Near-repeat modeling

Find a high risk of a violent
gang/criminal intelligence (on recent intergroup
outbreak between criminal groups
reports violence)

Clinical tools that Regression and

Identify individuals who may
summarize known risk classification models
become offenders:
factors using the risk factors

Probationers and parolees at

greatest risk of reoffending

Here, conventional methods rely on clinical techniques that add up the number of risk
factors to create an overall risk score. The corresponding predictive analytics methods
use regression and classification models to associate the presence of risk factors with
a percent chance that a person will offend. Also of interest are methods that identify
criminal groups (especially gangs) that are likely to carry out violent assaults on each
other in the near future. Hence, these methods can also be used to assess the risk that
an individual will become a victim of crime”.

54
 Domestic violence cases with
a high risk of injury or death

Mental health patients at

greatest risk of future criminal
behavior or violence

Table 2 – Methods for predicting the future aggressor. Source: Predictive Policing.
The Role of Crime Forecasting in Law Enforcement Operations, in RAND. Safety and
Justice Program, 2013, 10.

4.3. Methods for drawing up a criminal identikit

The third category, however, includes the “methods for

drawing up a criminal identikit” (see Table 386). In a certain sense,
these models come under the nuance of the term “predictive” which
concerns the merely investigative stage for deeds already committed.
The objective of this type of strategies is the processing of data for
suspects using a multi-level database. These databases contain an
indefinite plurality of information that is completely diverse
(biometric data, records for the ownership of assets, tax information,

86
“Table 1.3 summarizes methods used to identify likely perpetrators of past
crimes. These approaches are essentially real-world versions of the board game
Clue™: They use available information from crime scenes to link suspects to crimes,
both directly and by processes of elimination. In conventional approaches,
investigators and analysts do this largely by tracing these links manually, with
assistance from simple database queries (usually, the names, criminal records, and
other information known about the suspects). Predictive analytics automate the
linking, matching available “clues” to potential (and not previously identified)
suspects across very large data sets”.

55
etc.), the aggregation of which, by using delegated algorithms if
necessary, provides the operators with brilliant leads in an
investigation.

Conventional Crime
Predictive Analytics
Analysis (low to
Problem (large data demand and
moderate data demand
high complexity
and complexity)

Identify suspects using a Manually reviewing

Computer-assisted queries
victim’s criminal history or criminal intelligence
and analysis of intelligence
other partial data (e.g., plate reports and drawing
and other databases
number) inferences

Determine which crimes are Crime linking (use a table

part of a series (i.e., most to compare the attributes of Statistical modeling to
likely committed by the same crimes known to be in a perform crime linking
perpetrator) series with other crimes)

Locating areas both near Geographic profiling tools

Find a perpetrator’s most
and between crimes in a (to statistically infer most
likely anchor point
series likely points)

Find suspects using sensor

Computer-assisted queries
information around a crime Manual requests and
and analysis of sensor
scene (GPS tracking, license review of sensor data
databases
plate reader)

Table 3 – Methods for drawing up a criminal identikit. Source: Predictive Policing.

The Role of Crime Forecasting in Law Enforcement Operations, in RAND. Safety and
Justice Program, 2013, 11.

56
 4.4. Methods for predicting the future victim

Finally, the last category is an interesting “synthesis” of the

methods expounded so far (see Table 487). In fact, the “methods for
predicting the future victim” use the tools in the previous categories
with the objective, this time, of identifying the groups or individuals
at risk of victimisation.

Conventional Crime Analysis Predictive Analytics

Problem (low to moderate data (large data demand
demand and complexity) and high complexity

Identify groups likely to be Advanced models to

victims of various types of Crime mapping (identifying identify crime types by
crime (vulnerable crime type hot spots) hot spot; risk terrain
populations) analysis

Advanced crime-
Manually graphing or mapping mapping tools to
Identify people directly most frequent crime sites and generate crime locations
affected by at-risk locations identifying people most likely to and identify workers,
be at these locations residents, and others who
frequent these locations

87
“Table 1.4 summarizes methods to identify groups—and, in some cases,
individuals—who are likely to become victims of crime. These methods mirror those
used to predict where and when crimes will occur, as well as some of the methods
used to predict who is most likely to commit crimes. Predicting victims of crime
requires identifying at-risk groups and individuals—for example, groups associated
with various types of crime, individuals in proximity to at-risk locations, individuals
at risk of victimization, and individuals at risk of domestic violence”.

57
 Advanced data mining
Identify people at risk for Review of criminal records of techniques used on local
victimization (e.g., people individuals known to be and other accessible
engaged in high-risk engaged in repeated criminal crime databases to
criminal behavior) activity identify repeat offenders
at risk

Computer-assisted
database queries of
Manual review of domestic multiple databases to
Identify people at risk of disturbance incidents; people identify domestic and
domestic violence involved in such incidents are, other disturbances
by definition, at risk involving local residents
when in other
jurisdictions

Table 4 – Methods for predicting the future victim. Source: Predictive Policing. The
Role of Crime Forecasting in Law Enforcement Operations, in RAND. Safety and
Justice Program, 2013, 12.

5. Predictive policing and scientific analysis models

The current trends for the analytical models applied to

predictive policing are quite diverse depending on the functionalities
that they allow to be exploited. At a general level, there are two macro-
sectors into which artificial intelligence can be inserted to fill in the
gaps of subjectivity, which is often inadequate for providing answers
capable of taking into consideration all the social factors in play.
On one hand, one can see the predictive models aimed at
predicting the place, time and risk factors of future criminal events; on
the other, there are technologies used to relate the characteristics of the
different crimes committed (crime linking).

58
 Three different categories of models 88, which in some ways
are complimentary, can be identified in the context of the first sector.

5.1. Models for identifying criminal hot spots

A first set of technical/scientific methodologies was designed

for the identification of the places where the probability of a crime
being committed was highest within a certain time frame.
The most well known model belonging to that category is the
so-called hot spot analysis, designed, in the final analysis, for so-called
crime mapping. Hot spot analysis means the statistical tools applied
for predicting the areas where a certain crime can be much more
frequent. In fact, it is obvious that crime is not uniformly distributed
within a given area and this allows us to distinguish the areas with the
highest rates of delinquency (so-called “hot spots”) from those where
a certain crime is rarer in probabilistic terms The basic probability
assumption, essential for the operation of the mechanisms under
analysis, is that the crime will be committed again in the places where
it has already been perpetrated, according to an astute futuristic
interpretation of repetita iuvant.

88
The proposed classification was first put forward in Predictive Policing.
The Role of Crime Forecasting in Law Enforcement Operations, in RAND. Safety and
Justice Program, 2013, 19 et seq. (available in full at the following address:
https://www.rand.org/content/dam/rand/pubs/research_reports/RR200/RR233/RAN
D_RR233.pdf), and was then taken up by A. DI NICOLA – G. ESPA – S. BRESSAN –
M.M. DICKSON – A. NICOLAMARINO, Metodi statistici per la predizione della
criminalità, cit., 19.

59
 “Hot spots” can be analysed using different technologies,
among which we will mention at this time: grid mapping, covering
ellipses, kernel density estimation and heuristic methods.
Grid mapping allows the production of a grid drawn up from
traditional Cartesian coordinates (see Figure 2). The area of a certain
territory is surveyed into single cells, in each of which the quantity of
crimes committed is measured89. The operation described leads to
highlighting certain cells (so-called “hot”) where the risk of a crime
being committed appears high. Despite the undoubted usefulness of
maps drawn up like that, it is not possible to state with certainty
whether the isolated cells denote areas actually at risk or, whether, on
the contrary, they represent the results of completely exceptional
events.

89
The maps and surveying into single cells are processed using known GIS
(Geographical Information System) analysis systems. See S. CHAINEY – L. TOMPSON,
Engagement, Empowerment and Transparency: Publishing Crime Statistics using
Online Crime Mapping, in Policing, 2012.

60
 Figura 2 – Grid mapping of the thefts in Washington, D.C.
Fonte: Preliminary findings from an RTI International,
Structured Decisions Corporation, and RAND project funded
by NIJ’s Office of Science and Technology (OST).

On the other hand, covering ellipses exploit the product of the

map to aggregate and correlate the different contiguous hot spots.
These are grouped into clusters and end up being enclosed inside
various ellipses (see Figure 3). Not all the areas enclosed inside an
ellipse will be hot spots; there is specific software capable of removing
the areas that are not of interest from the area enclosed in the curve
(so-called two-step approach). An example of this specific technique
used to aggregate hot spots is “nearest neighbour hierarchical
clustering” (NNH), which identifies the points of interest based on pre-
determined criteria. Unlike the two-step approach, the elliptical
clustering methods presents (at least) two significant criticisms.

61
Firstly, it requires the numbers of ellipses to be predetermined in
advance. Since predetermining the number is often impracticable, the
problem is tendentially overcome by using a random choice of
ellipses, which are then tested to obtain the most accurate result. The
second problem, however, regards the results of the observation on the
area, which risks being highly influenced by the input data. In this
sense, for example, the results can be distorted depending on the time
interval in which the experiments were carried out: the width of the
margin of error is inversely proportional to the temporal extent of the
period of the study.

62
 Figura 3 – S.C. S MITH – C.W. B RU CE , CrimeStat III
User Workbook, in National Institute of Justice,
2010, 56.

“Kernel density estimation” (KDE) is another of the

approaches that allow hot spots to be identified. The kernel methods
are mathematical algorithms used, in this case, to identify hot spots on
the basis of the causal contributions of past crimes on the risk of future
crimes. The mathematical function called “kernel” models the data on
the incidence of crimes and, taking a determined spatial interval into
consideration, allows the values obtained to be positioned in the

63
corresponding geographical points90. This way, based on the
localisation of the crimes committed, an increasing quantity of data
can be mapped (see Figure 4). If the hot spots are created solely on the
“already committed crimes” variable, this is called “single kernel
density estimation”. However, if the variable concerning the
population density of the area is also used, this called “dual kernel
density estimation”.

Figura 4 – Confrontation between the single and the dual Kernel density
estimation representing the aggressions in Colonia (Germany) in 2012.
Source: D. O BE RWITT LER – M. WIESEN HÜTTER , The Risk of Violent
Incidents Relative to Population Density in Cologne Using the Dual
Kernel Density Routine, in Ned Levine and Associates, 2004, 8 ss.

Finally, heuristic methods allow the hot spots to be identified

based on the experience and personal and professional capacities of
the law and order operators (hence the attribute “heuristic”). The
manual identification of hot spots on pin maps comes under these
methods. This works through the manual noting, by tasked officers, of
the crimes committed on the maps of the area: “quadrat thematic
mapping” (see Figure 5), which is the analogue version of grid maps;

90
T.J. SULLIVAN – W.L. PERRY, Identifying Indicators of Chemical,
Biological, Radiological, and Nuclear (CBRN) Weapons Development Activity in
Sub-National Terrorist Groups, in Journal of the Operational Research Society, 2004.

64
the jurisdiction bounded areas method (see Figure 6), which adds,
compared to the previous instrument, the geographical information for
the boundaries in force in the area (division into regions, districts, post
codes, etc.).

Figure 5 – Quadrat thematic

mapping for motor vehicle thefts.
Source: J. ECK – S. CHAINEY – J.
CAMERON – M. LEITNER – R.
WILSON, Mapping Crime, cit., 25.

Figura 6 – Jurisdictional-
bounded areas mapped having
regards to thefts of motor
vehicles. Source: J. E CK – S.
C HAINEY – J. C AMERON – M.
LEITNER – R. W ILSON ,
Mapping Crime, cit., 24.

65
 In the same category of analysis models, together with the
identification of hot spots, there are the “regression methods”. An
estimate of the risk of a crime being committed in the future is made
using these regression methods. The tools under analysis exploit the
interrelation existing between different types of variables: on one
hand, the danger that a crime will be committed and, on the other, a
vast range of further information regarding past crimes (see Figure 7).

Figura 7 – Multiple regression model having

regards to the robberies in Washington, D.C.
Source: Preliminary findings from an RTI
International, Structured Decisions
Corporation, and RAND project funded by NIJ’s
Office of Science and Technology (OST).

66
 A further technique for constructing mathematical models,
again belonging to the first category and aimed at predicting future
crimes, is so-called “data mining”, also called “predictive analytics”.
At a practical level, a substantial set of computerized data is analysed
in the search for a pattern or trend that can be used for investigations.
The regression method, because it is mathematical, belongs to this type
of technique together with other methodologies such as classification,
and clustering tools and ensemble methods.
The aim of the classification methods is to predict a data
category based on an outcome. The clustering methods, on the other
hand, divide the data into groups on the basis of their similarities in
mathematical terms. Finally, the ensemble methods merge the use of
other simple predictive methods so that the final combined product can
supply more accurate and reliable results.
The last analysis method, belonging to the first category,
which is analysed here, is the “near repeat methods”. These techniques
start from the probability assumption that future crimes will be
committed in areas adjoining those where crimes were committed in
the past. An algorithm called “self exiting process” or “earthquake
modelling” has been created on the basis of similar theorising. Despite
it being a rather complex mathematical algorithm, the operating steps
that it follows are somewhat linear: a grid is represented on the
territorial map and, depending on the characteristics of the area, the
“background rate” which, when it is reached, it is probable that a new
crime may be committed, in the same way as with the prediction of
future earthquakes. It is taken for granted, in this process, that when a
crime arises, the repeat rate or that crime increases proportionally to
the passing of time (aftershock rate).

67
 5.2. Spatiotemporal analysis and geospatial risk factors

The second category of analysis models has the aim of

identifying the time when future crimes will most likely be committed.
Spatiotemporal analysis examines the evolution of crimes over time.
In order to complete that examination, it takes into consideration
diverse factors, among which the time of the day, the day of the week,
the temporal proximity to socially significant events (e.g. pay days,
cultural events, sporting events, etc.), the season and the weather. This
set of variables allows an effective prediction on both short-term
criminal series and long-term ones.
Finally, the third category is of particular use when you want
to identify the geospatial factors that increase the risk of a future crime
being committed. Using “risk terrain analysis”, which was developed
in the United States, statistical knowledge is exploited to identify the
geographical characteristics that increase the risk of crimes being
committed. The result, apparently, is similar to the one obtained
through the hot spot analysis techniques. Indeed, maps highlighting
the areas at risk are will be produced in both cases.
However, the difference is considerable: whilst in hot spot
maps the highlighted cells represent areas with the highest rate of past
crime, the future risks are highlighted in risk terrain analysis. This
distinctive feature is obviously a considerable strong point of the
methodology compared to the identification of hot spots.
Nevertheless, there are factors such as the area population that can
influence the analysis and make it unreliable: in a specific case, an area
with a high residential population risks being hot because of the
number of residents, when it actually presents no risk 91.

91
J.M. CAPLAN – L.W. KENNEDY (editor), Risk Terrain Modeling Manual:
Theoretical Framework and Technical Steps of Spatial Risk Assessment for Crime
Analysis, CreateSpace Independent Publishing Platform, 2010.

68
 6. The experience of predictive policing in Anglo-Saxon
countries

One can deduce from the analysis of the predictive policing

techniques that the greatest development of artificial intelligence is
this field has taken place in the Anglo-Saxon countries and, in
particular, the United States. The latter has developed several
predictive policing systems which, unlike the European experience
(see below for this), are used for wide range of crimes.
These predictive methods are designed to predict future
crimes in a perspective of general prevention. In fact, the American
software is not only based on the characteristics of specific crimes,
capable of providing the algorithms with the right data to predict future
actions (so-called crime linking), but, more generally, tends to analyse
the area and spreads the forces of law and order efficiently. With this,
they try to dissuade criminals, also thanks to the deterrence generated
in the associates, who are well aware of the operating efficiency of
artificial intelligence systems.
One of the more well known items of software in the
American scene is PredPol (abbreviation of predictive policing),
developed by researchers at two of the colleges in University of
California in collaboration with the area Police Department 92.

92
Z. FRIEND, Predictive Policing: Using Technology to Reduce Crime, in FBI
Law Enforcement Bulletin, 2013.

69
 Figura 8 – Heat map realized using the PredPol software. Source:
www.predpol.com.

The application, still in use in the United States, consists of an

operating diagram aimed at analysing property crimes, amongst
which, and in particular, burglary, vehicle theft and theft from
vehicles. The information on past crimes committed in the area under
analysis acts as input to the program, which already has a huge
database. This input is crossed with the software’s own algorithm,
which allows the “hottest” spots to be predicted. The heat map which
is traced (see Figure 8) is functional to the organisation of the police
operations, which, promptly and precisely, provide the system with
new data records, on the basis of which the analysts continue to
process trends and criminal models. It should be specified that the
police officers are not acquainted with the methods used to create the
maps: they are only required to use their professional skills and
experience to identify the interventions to be taken.

70
 One of the first models of predictive policing was born in 2005
in another area of the United States, more specifically in the city of
Memphis (Tennessee). The software, at the time innovative, is
developed by IBM with the name of Blue C.R.U.S.H. (Crime
Reduction Utilizing Statistical History). The city of Memphis, the
most populated in the whole of Tennessee, had seen a huge rise in the
crime rate, so much so as have significant effects on the population
and the local economy.
It is for this reason that the area authorities tried to draw up a
new methodology to combat and manage crime 93. The software project
took into consideration a set of variables, such as the geographical data
regarding the crimes committed and being committed, as well as
environmental, social and demographic factors. The crossing of the
data analysed allows hot spot maps to be created (see Figure 9). The
results obtained using this software were considerable and led directly
to a reduction of 25% in property crimes 94.

93
T. ARMSTRONG, Managing for 21st Century Crime Prevention in Memphis,
in www.managementexchange.com.
94
J. VLAHOS, Come anticipare il crimine, in Le scienze, 2012.

71
 Figura 9 – Mappa degli hot spot realizzata attraverso il software
Blue C.R.U.S.H. Fonte: IBM.

A similar experience, in Europe, is the one made in the United

Kingdom by researchers from the Jill Dando Institute of Security and
Crime Science in London. The experiment took place in Trafford with
the aim of drastically reducing the incidence of burglaries and was
Europe’s link with the experiences in the United States.
The project started from the assumption that repeat
victimization and near repeat victimization are repeated over time and
space, the reason for which it was necessary to spread police forces
over the areas where burglaries had been committed in the (not distant)
past. The operation of the experiments included an initial phase where,
one hand, the buildings in which the thefts had occurred were mapped
and, on the other, the buffer zones, highlighted with different colours

72
according to the near repeat victimization rate. The organisation of the
police, based on the data resulting from the experiment, allowed an
important reduction in burglaries to be achieved.

7. The experience of predictive policing in Europe

In Italy, the experience of artificial intelligence applied to

predictive policing is quite significant and is concentrated on the
prevention of specific crimes and the search for criminal profiles
which, sometimes, hide mandates that can be traced back to structured
criminal organisations capable of repeating (or have repeated) a
considerable quantity of crimes.
Besides the pioneering KeyCrime95, the experience of other
Italian born IT programmes is worthy of being mentioned, albeit
briefly, because it has led to satisfying results in the territorial areas of
interest.
One predictive policing piece of software that works through
a heuristic algorithm, on a probability basis, is XLAW. This instrument,
created by a Police Inspector, Elia Lombardo, is based on the idea that
urban crimes are committed in precise places and a relatively short
space of time, which allows the maximum profit to be drawn from the
seriality. The places where crimes are concentrated are chosen by
people acting on the basis of an objective element (e.g. presence of
potential victims) and a subjective element (e.g. presence of shelters,
suitability of the area in terms of accessibility and escape routes). The
software intelligence crosses the “appetising” places in a single map
and superimposes their socio-economic and environmental
characteristics, as well as information on past crimes, on it. This way,

95
For an in-depth analysis of this software, see L. Grossi, Chapter V, of this
research.

73
the system allows the re-creation of criminal models that can be
potentially applied to an indefinite series of crimes. The risk map
drawn up by the software is supplied to the operator, who, even two
hours in advance, can intercept the places and times where, at a
probability level, a crime will be committed.
Another predictive policing system, in the experimental stage,
is S.O.Cr.A.TE.S., software that is being researched by the Ministry of
Defence and the Department of Equal Opportunities, in collaboration
with the Department of Prison Administration, for scientific research
to be carried out with the support of inmates. The purpose of the
project is to construct effective criminal profiling that represents the
behaviours of various types of crime against the individual (violent
crimes, apparently motiveless and with a sexual background).
On the other hand, a computer program capable of receiving
anomalies that can be traced back to episodes of money-laundering is
being used in financial matters. Gianos is the software used by the
majority of Italian banking institutions: it is based on the constant
comparison of databases held by the various entities involved, which
allows a somewhat efficient cross-check to be carried out.
Predictive policing systems are also constantly used in
96
Spain . Already in 2006, following a series of forest fires in Galicia,
the Guardia Civil, in collaboration with the Fiscalía Coordinadora de
Medio Ambiente y Urbanismo de la Fiscalía General del Estado,
started a study on the psychological profile of so-called incendiarios
forestales. Police officers filled in an online questionnaire of a psycho-
social nature whenever they arrested a pyromaniac97. Thanks to the

96
See J.L. GONZÁLEZ ÁLVAREZ – J. SANTOS HERMOSO – M. CAMACHO
COLLADOS, Policía predictiva en España. Aplicación y retos futuros, in Behavior &
Law Journal, 6/2020, 26 et seq.
97
J.L. GONZÁLEZ – V. MUÑOZ – M.L. CALCERRADA – A. SOTOCA, Perfil
psicosocial del incendiario forestal español privado de libertad, in Behavior & Law
Journal, 3/2017, 26 et seq.

74
information obtained via the questionnaire instrument, predictive
policing tools were implemented. These allowed a search for the most
frequent characteristics in the pyromaniacs so as to facilitate the police
operators to localise and identify the criminal. The experiment was of
great use because it placed the methodological bases to be made use
of in other criminal dynamics.
The developments and application of predictive policing in
Spain are of particular significance in two specific sectors, which are
strictly connected to each other: on one hand, in gender violence and
management of the safety of the victims, and, on the other, in cases of
homicide, when the algorithms allow a probability estimate for the
personal characteristics of the potential authors to be made.
To date, the most developed predictive policing methodology
in Spain is the one regarding the creation and validation of a protocol
for assessing the risk of re-offending with regard to gender violence
using the Sistema de Seguimiento Integral de los Casos de Violencia
de Género (full monitoring system for cases of gender violence),
which is called VioGén and developed by the Secretaría de Estado de
Seguridad of the Ministry of the Interior. This protocol allows the
officers to assess the risk that a woman who reports a violent crime
runs of being a victim of the same crime from the same or another
partner. To do this, a specific computer procedure called Valoración
Policial del Riesgo (VPR: police risk assessment) is used98. Depending
on the level of risk obtained using the VPR, preventive strategies
aimed at anticipating the repeat criminal action are drawn up, with the
final aim of safeguarding the passive subject from against the repeat
victimization that he/she would be subject to99.

98
J.J. LÓPEZ OSSORIO – J.L. GONZÁLEZ – S. BUQUERÍN – L.F. GARCÍA – G.
BUELA CASAL, Risk Factors related to intimate partner violence police recidivism in
Spain, in International Journal of Clinical and Health Psychology, 2017, 107 et seq.
99
J.J. LÓPEZ OSSORIO – J.L. GONZÁLEZ ÁLVAREZ – J.M. MUÑOZ VICENTE – C.
URRUELA – A. ANDRÉS PUEYO, Validation and Calibration of the Spanish Police

75
 In the context of the fight against gender violence, actively
combated by the Spanish system (see Ley Orgánica no. 1/2004 on
global protective measures against gender violence), one should note
that the cases of so-called femicide in Spain reach absolute values that
cannot be ignored. It is for this reason that the Equipo Nacional de
Revisión Pormenorizada de Homicidios en el contexto de la Violencia
de Género (EHVdG) was set up in 2018, with the aim of examining
the case histories of femicide at international level as well as the best
practices of similar teams set up in other countries 100. The work group
was also engaged planning and promoting the monitoring of the cases
of femicide in Spain, facilitating the organisation of experts at district
level, which consisted of professionals from a large number of
universities and research institutes. These technicians, on the basis of
collaboration agreements with the Ministry of the Interior, undertook
the analysis work in the field, after receiving suitable training from the
national team.
One of the most significant results of the detailed study of the
femicides was the construction of a prediction scale for the risk of a
lethal outcome following the reporting of gender violence. An
analytical tracing of the crime facts was followed to create the scale,
called H. The final sample consisted of over two thousand criminal
episodes. Of these, a little less than a tenth resulted in femicides and
the remaining 90% consisted of non-mortal cases which had to be
constantly monitored101.

Intimate Partner Violence Risk Assessment System (VioGén), in Journal of Police and
Criminal Psychology, 2019.
100
J.L. GONZÁLEZ – M.J. GARRIDO – J.J. LÓPEZ OSSORIO – J.M. MUÑOZ – A.
ARRIBAS – P. CARBAJOSA – E. BALLANO, Revisión pormenorizada de homicidios de
mujeres en las relaciones de pareja en España, in Anuario de Psicología Jurídica,
2018, 28 et seq.
101
J.L. GONZÁLEZ ÁLVAREZ – J. SANTOS HERMOSO – M. CAMACHO COLLADOS,
Policía predictiva en España, cit., 33.

76
 The VPRs of all the episodes of crime about which there was
information were needed to complete the study, as the aim was to
understand whether the indicators for predicting a repeat of the
violence were capable of also predicting a fatal episode.
To make the public safety operators’ decisions easier, with
regard to the protection of the victims, it was also decided to program
a dual algorithmic mechanism that was transparent and shared with the
officers. This way, when a report of violence is received, the police
officers fill in the information card for the VPR. At this point, without
showing any type of result, the VioGén system applies the first
algorithm and calculates the risk of the specific episode being repeated
on the basis of the elements available at that time. Then, with
maximum speed, the program calculates, using the second algorithm,
the risk that episode could constitute only a part of the criminal design
that will culminate in femicide.
If the second algorithm gives a positive result (i.e. high risk of
a mortal event), the risk of re-offending as of the first algorithm is
raised a level. Only at this point is the result of the analysis using
artificial intelligence shown to the officers, with the warning that the
above-mentioned case is of particular interest. This takes place so that
the measures adopted by the police bodies can be adequate for the
characteristics and circumstances of the specific case.
One element of particular interest is that the report made to the
police officers is also recorded in special minutes, to be sent
immediately to the competent court and public prosecutor’s office.
The latter, on ascertaining the characteristics of the case, can adopt
suitable measures to safeguard the victim and, if necessary, provide
that the persons concerned are promptly assessed by psychologists or
medical examiners capable of going in depth into the factual
circumstances and proposing new or different measures to protect and
safeguard the victim.

77
 This dual mechanism is put forward to reduce the rate of
femicide in cases where a report has been made, clearly within the area
in the VioGén system is used. The real efficacy of this dual protocol
cannot be ascertained at the moment: in fact, the evolution of the
various crimes of violence over a significant time scale must be
analysed to monitor the work of the algorithms and formulate a
judgement on their work.

8. Criticisms, risks and legal limits: towards a “shared”

standardisation of predictive policing

Artificial intelligence, now at the height of its evolution and at

the centre of social and legal debate102, is an undoubtedly important

102
On this point, see “European Ethical Charter on the Use of Artificial
Intelligence in Judicial Systems and their Environment”, adopted by the European
Commission for the Efficiency of Justice (CEPEJ) during its 31st plenary meeting
(Strasbourg, 3-4th December 2018), 35, § 7: “Instruments described as “predictive
policing” (before the judicial process or before a court referral) are already growing
rapidly and are beginning to be known by the general public (for example, think of
the no fly list, which is actually a big data analytics application that collects and
analyses data on potential terrorists in order to prevent the commission of acts, or
algorithms used to detect fraud or money laundering). In general, a large number of
computer tools are commonly used to prevent the commission of criminal acts (by
identifying possible places where this could happen or their authors) or prosecute them
more effectively. The first category includes “predictive policing” tools that are used
to prevent certain types of offences with elements of regularity in their occurrence
such as burglary, street violence, theft from/of vehicles. The designation of these tools
derives from their ability to determine precisely where and when these offences could
be committed and to reproduce this information on a geographical map in the form of
hot spots that are monitored in real time by police patrols. This process is called
“predictive criminal mapping”. Most of the software used in this area is based on
historical crime location evidence, such as police reports, but even more powerful new
technologies combining various data and from different sources are also being tested.
These tools, which have very persuasive rates of effectiveness, are also claimed to

78
innovation which must, nevertheless, face its structural criticisms and
the legal limits set by the current system.
Firstly, it has been highlighted several times that the operation
of artificial intelligence systems (or rather, rationality) is based on the
collection and processing of big data. The quantity of information and
data collected is nothing but the fruit of reprocessing by the human
being, who, as such, incorporates various partiality factors. At an
instrumental level, this circumstance represents an effective limit to
the operations of the algorithms: in fact, they only produce effective
results if the input supplied is of quality and corresponds to the factual
reality.
In order to achieve satisfactory results, it would perhaps be
opportune to share and approve a numerus clausus of indispensable
variables to be considered so as to make the algorithmic procedure
much more objective. This closed number would only represent the
starting point for data collection. In fact, it is evident that, given the
specific circumstances of every case and the personal and professional
qualities of the operators, every datum collected could hide aspects
which escape the imperativeness of an “a priori” list.
A further problem encountered with data exploited by
algorithms regards their quantitative aspect. Algorithms struggle to
operate with the same efficiency in large and small situations. The
volume of data held by the operators, even if it were nothing but the
different crime rate (from which the information “useful” for
predicting the future is taken), diverges in a manner that cannot be
ignored, depending on whether the reference territory is a medium-
large size or medium-small size area.

have deterrent effects on the commission of offences in areas surrounding hot spots,
leading to a positive opinion of public policies.”

79
 Finally, with reference to the data, two more strictly legal
questions are worth raising, one on privacy and the other on the
ownership and control of the databases and the respective algorithms.
The implications that concern the right of privacy, with
reference to the data used by artificial intelligence algorithms, come
under the more general framework of the socio-technological
development of contemporary society, which has led, in a short time,
to a significant and sharp growth in the daily exchange of information,
both at a domestic and an international level. However, as happens in
the case of the big data supplied to the software in question, personal
data, and particularly sensitive data, is circulated in remote-controlled
mod and, consequently does not stop at the thresholds of state
boundaries. Therefore, a regulatory intervention directly at
supranational level became necessary.
Personal data are thus widely protected both by domestic
regulations and European ones. All the characteristics of the crimes,
which, for one reason or another, concern individuals (criminals,
victims or third parties) and are analysed by operators who feed the
artificial intelligence, come into this category. The European
legislator, with Regulation no. 679/2016 (General Data Protection
Regulation), has also intended to prepare a special statute for particular
categories of personal data, among which genetic and biometric data
and data “regarding health” stand out. In the perspective of the
configuration of the aforesaid statute, the range of the expression,
designed to be broad, “data regarding health” is of particular
importance. This locution, like a broad genus, includes the both the
following species of data: firstly, the so-called immediately sensitive
data (or sensitive in the strict sense) and, secondly, the so-called
indirectly sensitive data (or sensitive in the broad sense).
Personal data which, by their nature, are born and exhausted
in the phenomenological description of the individual’s state of health

80
come into the first category. The genetic characteristics of the
individual, the blood group and audiometric results are, for example,
immediately sensitive information. On the other hand, personal data
which, though appearing generic, hide intimate profiles are defined as
indirectly sensitive – and are in any case worthy of particular
protection. For example, the occasional or habitual use of substances
that create dependency (e.g. tobacco, alcohol and drugs) and the
common prescription for glasses are information which, once
acquired, reveal obvious clues on the state of health of the data subject.
Article 9 of the GDPR prohibits every type of processing for
some sensitive data, except if “processing is necessary for the
establishment, exercise or defence of legal claims or whenever courts
are acting in their judicial capacity” (letter f) and “processing is
necessary for reasons of substantial public interest, on the basis of
Union or Member State law which shall be proportionate to the aim
pursued, respect the essence of the right to data protection and provide
for suitable and specific measures to safeguard the fundamental rights
and the interests of the data subject” letter g).
Even if it may seem taken for granted, the algorithm organism
is fed with every type of data (sensitive or not), even the most intimate
(e.g. concerning health and criminal convictions 103) or apparently

103
In this regard, see Article 10 of the GDPR, titled “Processing of personal
data relating to criminal convictions and offences”: “Processing of personal data
relating to criminal convictions and offences or related security measures based on
Article 6(1) shall be carried out only under the control of official authority or when
the processing is authorised by Union or Member State law providing for appropriate
safeguards for the rights and freedoms of data subjects. Any comprehensive register
of criminal convictions shall be kept only under the control of official authority”. With
reference to Italian legislation on privacy, see also Article 2-octies of Legislative
Decree no. 196/2003 (so-called Privacy Code), titled “Principles for the processing of
data relating to criminal convictions and offences”: “1. With the exception of the
provisions of Legislative Decree no. 51 of 18th May 2018, the processing of personal
data relating to criminal convictions and offences or related security measures based
on Article 6(1) of the Regulation, which is not under the control of public authorities,

81
is allowed, under Article 10 of that Regulation, only if the processing is authorised by
Union or Member State law providing for appropriate safeguards for the rights and
freedoms of data subjects. 2. In the absence of the aforesaid provisions of law or
regulations, the processing of the data as of paragraph 1 and the guarantees as of the
same paragraph is identified with the Decree of the Ministry of Justice, to be adopt,
under Article 17, paragraph 3, of Law no. 400 dated 23rd August 1988, after hearing
the opinion of the Guarantor. 3. Without prejudice to paragraphs 1 and 2, the
processing of personal data relating to criminal convictions and offences or related
security measures is permitted if authorised by a regulation of law or, in the cases
provided for by the law, regulations regarding, in particular: a) the fulfilment of
obligations and the exercise of rights by the owner or data subject on labour law or in
any case in the context of employment, within the limits established by law,
regulations and collective contracts, pursuant to Articles 9, paragraph 2, letter b), and
88 of the regulation; b) the fulfilment of the obligations provided for by the law and
regulations on mediation aimed at the settling of civil and commercial disputes; c) the
verification or ascertainment of the requirements of good standing requirements,
subjective requirements or disqualifying pre-requisites in the cases provided for by
the law or regulations; d) the ascertainment of responsibility relating to claims or
events regarding human life, as well as the prevention, ascertainment and combating
of fraud or situations of concrete risk to insurance activities, within the limits provided
for by the law or regulations on the subject; e) the ascertainment, exercise and defence
of a right in court; f) the exercise of the right to access data and administrative
documents, within the limits provided for by the law or regulations on the subject; g)
the execution of investigations or searches or collection of information on behalf of
third parties pursuant to Article 134 of the Consolidated Text of the Public Safety
Laws; h) the fulfilment of the obligations provided for by the laws on anti-Mafia
communications and information or on the prevention of Mafia type delinquency and
other serious forms of social danger, in the cases provided for by the law and
regulations, or to submit the documentation prescribed by law to participate in
invitations to tender; i) the ascertainment of the moral fitness of those that intend to
participate in invitations to tender, in fulfilment of the provisions of current
regulations on invitations to tender l) the implementation of the discipline on
attributing the legality rating for companies under Article 5-ter of Act no. 1 of 24th
January 2012, converted with amendments by Act n. 27 of 24th March 2012; m) the
fulfilment of the obligations provided for by current regulations on preventing the
financial system from being used for the laundering of revenue obtained from criminal
activities and the financing of terrorism. 4. If the provisions as of paragraph 3 do not
identify the appropriate guarantees for the rights and freedoms of the data subjects,
such guarantees are provided for by the decree as of paragraph 2. 5. When the
processing of the data as of this Article is under the control of public authorities, the
provisions as of Article 2-sexies are applied. 6. The processing of the data as of Article
10 of the Regulation is authorised with the decree in article 2. This carried out in
implementation of the protocols of understanding for the prevention and combating

82
useless data. Currently, the current regulations on privacy applied to
computer databases could be inadequate or, in any case, non-
exhaustive whilst the ownership of the personal data, the subject of the
processing, remains in the hands of private entities, with a heavy
limitation of public control.
The artificial intelligence algorithms studied so far and the
software in use are, indeed, under the almost total hegemony of private
entities. The companies or organisations which create and own the
algorithms exploit, almost freely, the personal data of a very high
number of individuals.
On this point, it is interesting to note that these algorithmic
codes seem inaccessible to the public: a lack of transparency is hidden
behind this context, which, normally, belongs to the action of the state
administration104. In this regard, it is interesting to mention the recent
ruling of the Lazio T.A.R. [Regional Administrative Court], sect. III
bis, n. 3769/2017, which, although in a different sector and in different
circumstances, recognised the right of the petitioner, as the data
subject, to access the algorithm used by the public administration in
managing the proceedings under its remit, based on the fact that the
very algorithm, all in all, gives life to the administrative act 105.

of organised crime, which were drawn up with the Ministry of the Interior and the
Prefectures-UTG [Government Territorial Offices]. With regard to these protocols,
the decree, as of paragraph 2, identifies the types of data processed, the data subjects
and the processing operations that can be carried out, also in relation to updating and
storage, and provides the appropriate guarantees for the rights and freedoms of the
data subjects. The decree is implemented, limited to the contexts as of this paragraph,
in agreement with the Ministry of the Interior”.
104
On the problem of the transparency of the algorithms, see E. GABELLINI ,
La «comodità nel giudicare»: la decisione robotica, in Riv. trim. dir. proc. civ.,
4/2019, 1305 et seq.
105
The sentence under analysis recognised the right of the petitioner because
the algorithm, though being pre-set to execute the duties established by school
regulations, finished up by deciding de facto what should be the destination of the
teacher when the scholastic institutes of service are assigned. In this sense, the

83
 In conclusion, no critical issues seem to arise relating to the
usefulness of the evidence as far as artificial intelligence systems
applied to predictive policing are concerned. With the exception of the
privacy problem, the whole process of “predicting the future” does not
seem to present any significant problems as long as the computer
systems are solely an aid to the operators so that they can have a more
efficient organisation, as happens in the case of preparing heat maps.
However, it is clear that if the “artificial” systems are
functional to the establishment of evidence to be used in penal action,
the permissiveness of the current system is less clear and leaves space
for wide profiles of uncertainty106.
In the context of possible co-ordination between predictive
policing and predictive justice artificial intelligences, it would seem
desirable to create an algorithmic system that allows you to reconstruct
the elements making up the fact to be created for each macro-category
of crimes, as much from an objective perspective as a subjective one.
If the investigators could enter all the investigative elements available
into a computer system and this system could manage to “pre-analyse”
them in the framework of an informal “pretrial”, the debating dossier
would benefit from it and the trial proceedings would meet less
obstacles, both in qualitative terms and with regard to the procedural
economy required by due process.

algorithmic process is an integral part of the administrative process which gives life
to the consequent act and, therefore, comes under access to acts governed by Act no.
241/1990. On this point, see I. FORGIONE, Il caso dell’accesso al software MIUR per
l’assegnazione dei docenti – T.A.R. Lazio Sez. III bis, 14 February 2017, n. 3769, in
Giornale di diritto amministrativo, 2018, 647 et seq.; L. VIOLA, L’intelligenza
artificiale nel procedimento e nel processo amministrativo: lo stato dell’arte, in Foro
amm., 2018, 1598 et seq..; A. SIMONCINI, L’algoritmo incostituzionale: intelligenza
artificiale e il futuro delle libertà, in BioLaw Journal, 1/2019, 73 et seq.
106
On the subject of the legal limits to the usefulness of the evidence, see
below, L. NOTARO, Chapter III of this research.

84
 In conclusion of this analysis, one cannot omit the hope that,
in the perspective of an implementation of artificial intelligence tools
in the near future, a regulatory proposal is put forward, also at
European level (with the collaboration of the bodies involved, such as
Europol and EPPO), and aimed at regulating the automated crime
prevention (in the broad sense) processes uniformly.
In order to do this, it undoubtedly becomes necessary to share
knowledge, as the knowledge required to meet the transition from
traditional systems to “artificial” ones, so to speak, involve a diverse
multitude of disciplinary sectors, ranging from mathematics, computer
science and statistics to law107.

107
On the opening of law to other sciences and the need to share knowledge,
see P. GARBOLINO, Nuovi strumenti logici e informatici per il ragionamento
giudiziario: le reti bayesiane, in Cass. pen., 1/2007, 326 et seq.

85
 CHAPTER III

ARTIFICIAL INTELLIGENCE AND CRIMINAL JUSTICE

Laura Notaro

Table of Contents: 1. Experiences of application. – 1.1. The United States experience.

– 1.1.1. Decisions on pretrial release. – 1.1.2. Sentencing. – 1.2. European
tendencies and experiences. – 2. Possible fields of application for AI in
criminal justice: a systematic overview. – 2.1. Assessment of the probability
of repeat offending. – 2.2. Prediction of the decision-making result. – 2.3.
Further theoretical fields of application. – 3. Criticisms and the risks of using
AI technologies in the criminal system. – 3.1. Risk-assessment tools. – 3.1.1.
Some observations. – 3.2. Tools for predicting the decision-making result. –
3.3. Other fields of application. – 4. Legal limits on the entry of AI into the
criminal system. – 4.1. The Ethical Charter of the CEPEJ of the Council of
Europe. – 4.2. European Union law: the limits to “decisions based solely on
automated processing”. – 4.3. The limits of domestic law: the ban on
criminological expert opinions. – 5. Final remarks. Spaces of application and
precautions for AI in criminal matters.

1. Experiences of application

The use of artificial intelligence tools in the context of

criminal justice has also become an ever more pressing subject in
Europe.
It is true that the most significant experience of applying
technologies in the criminal court is still today the one offered by the
United States system. Nevertheless – and perhaps precisely because of
this – initiating a reflection at European level is needed at this point.
Formulating the theoretical debate on the nature of these tools and the

86
compatibility of their application in criminal justice systems with the
fundamental rights guaranteed at European level and the constitutional
principles of the member states would allow us to precede the
development of technology and practices and thus prepare a solid
cultural base108 and a clear picture of legal regulations, fit for facing
the challenges of modernity.
The survey proposed here can be broken down into five
sections.
The first and second part will define the outlines of the
research: firstly, the main experiences of using AI tools in criminal
justice in the United States and Europe will be examined and,
secondly, an attempt will be made to provide a systematic framework
for the possible fields of application for artificial intelligence in the
context of criminal justice.
The third and fourth part will continue with an analysis of the
problems arising in relation to the fundamental rights guaranteed at
European level, the general principles of law and the criminal trial and
the further juridical limits, of a supranational and domestic source, to
the entry of AI technologies in the criminal justice system.
The final part will make an attempt to define the places of
application for AI in the criminal system in light of the results of the
previous phases of investigation, also advancing the necessary
precautions so that the use of these technologies is compatible with the
national and European picture.

108
A warning on the need that the execution of public policies for the
implementation of AI tools must be preceded by a multi-discipline public debate that
delineates a picture of algorithm development in line with compliance of the
fundamental rights is given by X. RONSIN, V. LAMPOS, V. MAÎTREPIERRE, In-depth
study on the use of AI in judicial systems, notably AI applications processing judicial
decisions and data, in Appendix I, European ethical Charter on the use of Artificial
Intelligence in judicial systems and their environment, in www.coe.int.

87
 1.1. The United States experience

The predictive software popular in the United States is the

most representative of the tendency to introduce AI systems into
criminal justice, since it is the first and most advanced experience in
this direction.
There are two main spheres of application: a) the decisions on
pretrial release; b) the sentencing stage.
Although the computer tools used are very numerous, the most
well-known and widespread are “COMPAS” (Correctional Offender
Management Profiling for Alternative Sanctions), “LSI-R” (Level of
Service Inventory Revised), both developed by commercial
companies, and “PSA” (Public Safety Assessment), created by the
Arnold Foundation109. Some states however have developed their own
predictive algorithms 110.

109
For a recognition of the predictive tools used in each state, see the
document published by EPIC (Electronic Privacy Information Centre), Algorithms in
the Criminal Justice System: Risk Assessment Tools, available at the following link:
https://epic.org/algorithmic-transparency/crim-justice/. On this subject, D. KEHL, P.
GUO, S. KESSLER, Algorithms in the Criminal Justice System: Assessing the Use of
Risk Assessments in Sentencing (Responsive Communities Initiative, Berkman Klein
Center for Internet & Society, Harvard Law School), July 2017, available at the
following link: https://cyber.harvard.edu/publications/2017/07/Algorithms.
110
For example, the Ohio Department of Rehabilitation and Correction has
developed the algorithm “ORAS” (Ohio Risk Assessment System) in collaboration
with the University of Cincinnati (in this regard, D. KEHL, P. GUO, S. KESSLER,
Algorithms in the Criminal Justice System, cit., 16). On the algorithm recently adopted
in Pennsylvania, as the result of a work started in 2010 by the Pennsylvania
Commission on Sentencing, see R. HESTER, Risk Assessment at Sentencing. The
Pennsylvania Experience, in VARIOUS AUTHORS, Predictive Sentencing: Normative
and Empirical Perspectives, edited by J.W. de Keijser, J.V. Roberts, J. Ryberg,
Bloomsbury Publishing, 2019, 213 et seq.; A. DIENER, Pennsylvania’s Proposed,
Questionably Constitutional, Risk Assessment Instrument, in Harvard Civil Rights –
Civil Liberties Law Review, 17th October 2019; A. BASHIR, Pennsylvania’s Misguided
Sentencing Risk-Assessment Reform, in The Regulatory Review, 5th November 2019.

88
 However, it is appropriate to specify that the spread of
computer systems for risk analysis in the criminal justice system of the
United States comes under the more general and dated tendency of
making use, in this context, of predictive tools constructed on the basis
of statistical-actuarial theories and methods 111. As highlighted by
careful doctrine112, the scientific theory translated into the algorithm
and the “computational” aspect of the instrument can in fact be
distinguished113, at a conceptual level, so that the different problems
arising from each of the two characteristics can be analysed properly.

111
On the affirmation of evidence-based practices (EBP), see D. KEHL, P.
GUO, S. KESSLER, Algorithms in the Criminal Justice System, cit., 7. In this regard,
some criticisms are highlighted by The case against categorical risk estimates, in
Behavioral Science Law, 2018, 1 et seq.
112
S. QUATTROCOLO, Equo processo penale e sfide della società algoritmica,
in Riv. BioDiritto, 1/2019, 135 et seq.; EAD., Questioni nuove e soluzioni antiche?
Consolidati paradigmi normativi vs rischi e paure della giustizia digitale
“predittiva”, in Cass. pen., 4/2019, 1478 et seq.
113
In this sense, S. QUATTROCOLO, Equo processo penale e sfide della società
algoritmica, cit., 144. For example, the predictive system “SAVRY” (Structured
Assessment of Violence Risk in Youth), used in juvenile justice by at least 9 states, is
a tool with an algorithmic structure, but without a digital/computational guise. In fact,
it is not software, but a user manual accompanied by assessment cards. On this subject,
see G.M. VINCENT, J. CHAPMAN, N.E. COOK, Risk-Needs Assessment in Juvenile
Justice: Predictive Validity of the SAVRY, Racial Differences, and the Contribution of
Needs Factors, in Criminal Justice & Behavior, 2011, 47 et seq.; S. QUATTROCOLO,
Questioni nuove e soluzioni antiche? Consolidati paradigmi normativi vs rischi e
paure della giustizia digitale “predittiva”, in Cass. pen., 4/2019, 1478 et seq.
For a broad notion of algorithm, see T. GILLESPIE, The relevance of
Algorithms, in T. GILLESPIE, P. BOCZKOWSKI, K. FOOT, Media Technologies, MIT
Press, 2014, 1: “Algorithms need not be software: in the broadest sense, they are
encoded procedures for transforming input data into a desired output, based on
specified calculations”. See also the study of the Council of Europe, Algorithms and
Human Rights, DGI (2017) 12, 5, in www.coe.it, 22nd March 2018, 5.

89
 1.1.1. Decisions on pretrial release

As regards the subject of pretrial release, there are more than

20 different risk assessment tools (RATs) in use in the United States
courts114. The consideration of these tools is required by law, at least
in certain cases, in 7 states (Alaska, Delaware, Hawaii, Indiana,
Kentucky, New Jersey and Vermont), whilst their adoption is
permitted and encouraged in another 8 states (Colorado, Illinois,
Montana, New York, Pennsylvania, Rhode Island, Virginia and West
Virginia)115. However, it should be noted that some legislative
interventions at state level (Idaho, New York and California) have
introduced regulations on the use of predictive tools in pretrial release
decisions: requirements of impartiality and non-discrimination,
guarantees of transparency and accessibility of the systems and the
data at the base of the algorithms, as well as the provision of periodic
assessment of the predictive instrument 116.
When examining the United States experience, it should be
borne in mind that the use of risk assessment tools in the context of
pretrial decisions must be inserted in a legal system that provides for
forms of money bail as a condition for freeing the arrested person
awaiting trial117. In this context, the introduction of algorithmic risk
assessment tools must be strictly tied to the debate on the abolition of

114
A. WIDGERY, The Statutory Framework of Pretrial Release, in
www.ncsl.org, 8th November 2020, 7. On this subject, A.Z. HUQ, Racial Equity in
Algorithmic Criminal Justice, in Duke Law Journal, 2019, 1043 et seq. In the Italian
doctrine, see M. GIALUZ, Quando la giustizia penale incontra l’intelligenza
artificiale: luci e ombre dei risk assessment tools tra Stati Uniti ed Europa, in Dir.
pen. cont., 29th May 2019, 4.
115
A. WIDGERY, The Statutory Framework of Pretrial Release, cit., 7.
116
In 2019, one should note House Bill 118 in Idaho, Senate Bill 1509 in the
State of New York and Senate Bill 36 in California.
117
On the subject of bail in the Italian doctrine, see V. TONDI, Il Bail. La
libertà su cauzione negli ordinamenti anglosassoni, Cedam-Wolters Kluwer, 2016.

90
bail118. The affair of California is representative of the question, also
at political level. The State of California was, in fact, the only one to
go ahead with the complete abolition of money bail in 2018, replacing
it with the obligation to make use of risk assessment tools to assess
whether and on what conditions bail should be granted. Nevertheless,
the reform was repealed in 2020 following the referendum of 3rd
November, supported by a diverse front, comprising, on one side, the
insurance companies interested in keeping bail money (the so-called
bail industry) and, on the other, some human rights organisations
which, though being favourable to the abolition of bail, had been
indicating the risk of machine bias and the poor transparency of
predictive software.
That said, among the algorithms most used in the context of
pretrial decisions, one should mention “PSA” (Public Safety
Assessment)119. This predictive instrument is developed by the “Laura
and John Arnold Foundation” non-profit organisation and is used in 4
states (Arizona, Kentucky, New Jersey and Utah) and in numerous
important jurisdictions in the United States120.
The input algorithm consists of 9 risk factors for the
individual121, which are compared with a database of about 750,000

118
M. GIALUZ, Quando la giustizia penale incontra l’intelligenza artificiale,
cit., 8; J.L. KOEPKE, D.G. ROBINSON, Danger Ahead: Risk Assessment and the Future
of Bail Reform, in Washington Law Review, 4/2018, 1725 et seq.
119
In this regard, in the Italian doctrine, see M. GIALUZ, Quando la giustizia
penale incontra l’intelligenza artificiale, cit., 7; F. BASILE, Intelligenza artificiale e
diritto penale: quattro possibili percorsi di indagine, in Dir. pen. uomo, 29th
September 2019, 18.
120
Among these, Allegheny County (Pittsburgh, Pennsylvania), Cook County
(Chicago, Illinois), Harris County (Houston, Texas), Mecklenburg County (Charlotte,
North Carolina), Milwaukee County (Wisconsin) and San Francisco County
(California). See the card “Where is PSA currently used?” in advancingpretrial.org.
121
The elements considered are: 1) age at current arrest; 2) current violent
offense; 3) pending charge at the time of the arrest; 4) prior misdemeanour conviction;
5) prior felony conviction; 6) prior violent conviction; 7) prior failure to appear in the

91
cases from about 300 jurisdictions.
The output returned by the system consists of allocating a
score of between 1 and 6 for each of the three “risks” to assess: Failure
To Appear – FTA, New Criminal Arrest – NCA and New Violent
Criminal Arrest – NVCA.
The criteria for allocating the scores are published online122.

1.1.2. Sentencing

More than 60 different risk assessment tools are used in the

sentencing stage123. The recourse to predictive tools is obligatory in
the legislation of some states124.
It is precisely the sentencing stage that the now well-known
Loomis case refers to, on which the Supreme Court of Wisconsin 125
pronounced, and in which the use of the “COMPAS” algorithm was
prominent 126.

past 2 years; 8) prior failure to appear older than 2 years; 9) prior sentence to
incarceration. See the card “How It Works”, in advancingpretrial.org.
122
See the card “How It Works”, in advancingpretrial.org.
123
A.Z. HUQ, Racial Equity in Algorithmic Criminal Justice, cit., 1075; A.M.
BARRY -JESTER, B. CASSELMAN, D. GOLDSTEIN, Should Prison Sentences Be Based on
Crimes That Haven’t Been Committed Yet?, in fivethirtyeight.com, 4th August 2015.
More generally, on the use of algorithms in sentencing, see D. KEHL, P. GUO, S.
KESSLER, Algorithms in the Criminal Justice System: Assessing the Use of Risk
Assessments in Sentencing, cit., 13 et seq.
124
D. KEHL, P. GUO, S. KESSLER, Algorithms in the Criminal Justice System:
Assessing the Use of Risk Assessments in Sentencing, cit., 16; A.Z. HUQ, Racial Equity
in Algorithmic Criminal Justice, cit., 1075. For example, the legislations of Arizona,
Kentucky, Ohio and Pennsylvania require the use of predictive tools in the sentencing
stage. Moreover, the last two states have developed their own risk assessment
algorithm.
125
State v. Loomis, 881 N.W.2d 749, 753 (Wis. 2016)
126
For an in-depth study, see below, L. GROSSI, Chapter IV of this research.

92
 It concerns a computational predictive instrument developed
in 1998 by the company Northpointe (now Equivant).
The database on which the operation of the system is based
consists of 300,000 cases examined by COMPAS between January
2004 and November 2005127.
The algorithm input is the information collected from the
accused’s file and the replies given by the subject to a series of
questions128.
The output returned by the system consists of a risk
assessment and a needs assessment. The part for risk assessment
allocates a score on a scale of 1 to 10 for each of the three measured
risks of recidivism: pretrial recidivism risk, general recidivism risk
and violent recidivism risk.
The picture of the factors considered and the results returned
by COMPAS is therefore very broken down and an indicator of a
datum that should not be overlooked: that the Northpointe algorithm
was not initially developed for the sentencing stage129, but to aid
judges and other operators in pretrial release decisions and during the
execution of the sentence (e.g. admission for parole).

127
Practitioner’s Guide to COMPAS Core (2019), in www.equivant.com, 11.
128
State v. Loomis, cit., § 13. The COMPAS algorithm can consider up to 137
factors (for this information, see X. RONSIN, V. LAMPOS, V. MAÎTREPIERRE, In-depth
study on the use of AI in judicial systems, notably AI applications processing judicial
decisions and data, in Appendix I, European ethical Charter on the use of Artificial
Intelligence in judicial systems and their environment, in www.coe.int, § 129; J.
DRESSEL, H. FARID, The accuracy, fairness, and limits of predicting recidivism, in
Science Advances, 4/2018, 1; J. NIEVA FENOLL, Intelligenza artificiale e processo,
translated by di P. Comoglio, Giappichelli, 2019, 59).
129
D. KEHL, P. GUO, S. KESSLER, Algorithms in the Criminal Justice System:
Assessing the Use of Risk Assessments in Sentencing, cit., 11.

93
 1.2. European tendencies and experiences

 HART (United Kingdom)

The technology that comes closest in Europe to the predictive

tools spread over the United States is “HART” (Harm Assessment
Risk Tool)130, under experimentation in the United Kingdom.
It concerns an algorithm developed through the collaboration
of the University of Cambridge and the Constabulary of Durham to
assess the risk of recidivism in the two years after the arrest. The
application was designed to be a better implementation of
Checkpoint131, an alternative intervention program to criminal
prosecution, aimed at the authors of crime, the risk of recidivism for
which concerns non-serious crimes. In this context, the algorithm has
the function of aiding the police in the selection of who the intended
recipients of the programme are.
HART, in use since 2017132, in fact collocates the individual
offenders into three risk categories regarding the risk of committing a
new crime in the two subsequent years: a) High risk (risk of serious
crimes being committed, such as homicide, attempted homicide,
grievous bodily harm, and sexual and firearms offences); b) Moderate

130
On this subject, M. OSWALD, J. GRACE, S. URWIN, G.C. BARNES,
Algorithmic risk assessment policing models: lessons from the Durham HART model
and “Experimental” proportionality, in Information and Communications
Technology Law, 2/2018, 223 et seq. In the Italian doctrine, M. GIALUZ, Quando la
giustizia penale incontra l’intelligenza artificiale, cit., 10.
131
Information on the Checkpoint project, active since 2015, is available on
the Durham Constabulary website (www.durham.police.uk). On this subject, see M.
OSWALD, J. GRACE, S. URWIN, G.C. BARNES, Algorithmic risk assessment policing
models, cit., 227.
132
S. CARLO, Big Brother Watch’s written evidence on algorithms in the
justice system for the Law Society’s Technology and the Law Policy Commission, in
bigbrotherwatch.org.uk, 2.

94
risk (risk of non-serious crimes being committed, mainly crimes
against property and minor offences regarding drugs); c) Low risk (no
risk of recidivism)133. The Checkpoint programme is reserved for the
authors of crime in the second category.
The HART input consists of 34 predictive factors: 29 are
connected with the individual’s criminal history; the others concern
the age, gender, two difference post codes and the number of police
intelligence reports on the author of the crime134. The algorithm is
constructed on a database of 104,000 cases in Durham between 2008
and 2012135.
A significant aspect is then the different weight given to
“dangerous errors (false negatives) and “cautious errors” (false
positives) when programming the algorithm: the choice in favour of
greater accuracy of the low risk assessments determines an
overestimate of high risk individuals 136. False negatives are thus very
rare, whilst one will have to take into account the possibility of getting
a not insignificant percentage of false positives.
One of the most widely discussed subjects in civil society, also
in light of the analyses conducted by non-profit organisations for the
protection the fundamental rights, concerns one of the two post codes
included in the predictive factors making up the HART input. This
concerns the Mosaic code, a geo-demographic tool developed and sold
by the Experian marketing company137, which profiles adults in the

133
M. OSWALD, J. GRACE, S. URWIN, G.C. BARNES, Algorithmic risk
assessment policing models, cit., 227.
134
M. OSWALD, J. GRACE, S. URWIN, G.C. BARNES, Algorithmic risk
assessment policing models, cit., 228.
135
Ibidem.
136
M. OSWALD, J. GRACE, S. URWIN, G.C. BARNES, Algorithmic risk
assessment policing models, cit., 230.
137
The product brochure can be consulted at the following link:
www.experian.co.uk/assets/marketing-services/brochures/mosaic-ps-brochure.pdf.

95
United Kingdom by putting them into 66 categories and works on a
base of 850 million data (amongst which, the composition of the
household and ethnicity and data regarding employment, health,
consumption, educational results and online data)138.

 Predictive justice commercial applications for

lawyers on civil and judicial analytics (France)

Although we are limiting ourselves to the context of civil

justice for the moment, some predictive tools developed in France are
worthy of mention.
Since 2016, when Loi No. 2016-1321 “sur la République
numérique” made all the judicial decisions available to the public,
some start-ups have used the very wide open access database to draw
up predictive algorithms designed to support the activities of
lawyers139.
The main service offered consists of indicating the probability
of success for an action and therefore aims to provide an aid for
lawyers in identifying the opportunity for undertaking an action and
the choice of the most appropriate strategy140.

138
On this point, see Big Brother Watch’s written evidence on algorithms in
the justice system, cit., 1 et seq.; BIG BROTHER WATCH TEAM, A closer look at
Experian Big Data and Artificial Intelligence in Durham Police, in
bigbrotherwatch.org.uk, 6th April 2018; BIG BROTHER WATCH TEAM, Police uses
Experian marketing data for AI custody decisions, in bigbrotherwatch.org.uk, 6th
April 2018.
139
There are at least four of them, developed by Predictice, Case Law
Analytics, Doctrine.fr, Tyr Legal. See C. SZWARC, La justice predictive: une autre
justice?, in Le Mag des Avocats, 34, 9/2017, 5 et seq., available at the following link:
https://www.anased.fr/publications/mag34/files/assets/basic-html/page-5.html#.
140
The French Ministry of Justice has started experimenting with one of these
algorithms (Predictice) at the Courts of Appeal in Rennes and Douai. See C.

96
 The algorithms developed thanks to the open access judicial
decisions that aroused the greatest perplexity were those that gave
importance to the identity of the individual judge in order to analyse
his/her work. In fact, the question did not emerge from the use of one
of commercial applications of the LegalTech start-ups, but from the
publication of a study on the SupraLegem site which provocatively
highlighted the considerable differences in the percentages of
dismissal for the different judges in the Conseil d’Etat who were
competent for appeals on asylum law141.
Foreseeing the risk that the profiling of an individual
magistrate could have had for the independence of the judiciary142 and
responding to the mobilisation of the trial judges, the French legislator,
in 2019, provided that the names of magistrates and data that could
identify them should be omitted from the material made available to
the public. A ban on every use of data identifying magistrates “with
the purpose of, or by effect, an assessment, analysis or comparison of
such data or to predict their future real or supposed directions”143 was
also introduced with criminal penalties.

CASTELLI, D. PIANA, Giustizia predittiva. La qualità della giustizia in due tempi, in

Quest. giust., 4/2018, 156.
141
The study was conducted by the lawyer Michaël Benesty and the IT
engineer Anthony Sypniewski. On this issue, see M. LANGFORD , M.R. MADSEN,
France Criminalises Research on Judges, in verfassungsblog.de, 22nd June 2019; M.
BENESTY, The Judge Statistical Data Ban. My Story, in www.artificiallawyer.com, 7th
June 2019.
142
B. GALGANI, Considerazioni sui “precedenti” dell’imputato e del giudice
al cospetto dell’IA nel processo penale, in Sist. pen., 4/2020, 87 et seq.
143
See article 33, al. 3, de la loi n. 2019-222 of 23rd March 2019: “Les données
d’identité des magistrats et des membres du greffe ne peuvent faire l’objet d’une
réutilisation ayant pour objet ou pour effet d’évaluer, d’analyser, de comparer ou de
prédire leurs pratiques professionnelles réelles ou supposées. La violation de cette
interdiction est punie des peines prévues aux articles 226-18,226-24 et 226-31 du code
pénal, sans préjudice des mesures et sanctions prévues par la loi n° 78-17 du 6 janvier
1978 relative à l’informatique, aux fichiers et aux libertés”.

97
  Toga (Italy)

With reference to Italy, the launch of a commercial application

called “Toga” should be mentioned 144. It offers the operators in the
criminal sector the possibility of consulting the whole current
legislation, immediately displaying all the institutes applicable to each
case of criminal act.
All the significant information for applicability to each crime
of the various substantial and procedural legal institutes (e.g.
recidivisms, non-punishability because of the particular minor nature
of the offence, plea-bargaining, etc.) is extracted from the database of
current laws by the algorithm.
The function of calculating the sentence is interesting. It
proposes technical support to the magistrates and lawyers which
prevents errors of calculation or the erroneous application of institutes
that are actually not provided for the crime in question.

 Experiments of predictive justice in civil matters: the

collaboration between the courts and university (Italy)

An experiment in predictive justice tools is also being started

up at some Italian courts. As in France, it concerns an experience
limited to the civil justice sector.
A first project was promoted by the Chairman of the Court of
Appeal of Brescia starting from 2018, in collaboration with the
Departments of Law and Statistics at the local university. It is aimed

144
The application is mentioned by A. TRAVERSI, Intelligenza artificiale
applicata alla giustizia: ci sarà un giudice robot?, in Quest. giust., 10th April 2019.

98
at creating a database of all the orders issued from 2018 onwards in
some subjects (enterprise, contracts, bank contracts, labour relations,
accidents and dismissals) and sees the participation of work groups
that extract the significant facts of the concrete case from the
individual orders 145.
A similar initiative was started by the Court of Genoa in 2019,
in collaboration with the DIRPOLIS institute of the Scuola Superiore
Sant’Anna of Pisa146.

 Experiments in systems of Online Dispute Resolution

and robot-judges in the civil context (Holland and Estonia)

Some further tendencies at European level in the civil context

must be mentioned for completeness.
An online legal aid platform has been active in Holland for a
while. This deals with separation and divorce and allows the spouses
to consensually plan the conditions for the dissolution of their
marriage147.
The initiative with the most invasive impact on justice is
potentially the one in Estonia, the Ministry of Justice of which started
the design in 2019 of software for deciding civil disputes with a value
of less than 7000 euros 148.

145
C. MORELLI, Giustizia predittiva: il progetto (concreto) della Corte
d’appello di Brescia, in www.altalex.it, 8th April 2019.
146
L’occhio della tecnologia sulla giustizia di Genova: un’intelligenza
artificiale potrà prevedere la sentenza?, in Sant’Anna Magazine, 5th October 2019.
147
The Rechtwijzer platform has been recently replaced by Justice42
(justice42.com). Other innovative justice projects in Holland concern platforms that
assist the parties in the joint consensual submission of their dispute to civil justice. On
this subject, see D. REYLING, Beyond court digitalization with ODR, in International
Journal for Court Administration, 8(2), 1.
148
J. PARK, Your Honor, AI, in Harvard International Review, 3rd April 2020.

99
 2. Possible fields of application for AI in criminal justice: a
systematic overview

Having briefly examined the current experiences of the use of

AI systems in the criminal justice context, it is convenient at this point
to go through the potential areas of significance for such tools, framing
them in the systematic “places” they belong to. The favoured
perspective will be the Italian system, though some very quick
comparative mentions will not be overlooked.

2.1. Assessment of the probability of repeat offending

As has been seen, the majority of AI systems experimented

with in criminal justice is applied to assessments for the probability of
re-offending, in its different spheres of relevance (in the context of bail
for the defendant during the trial or at the time of commutation).
If we want to briefly mention the institutes for the application
of which the Italian criminal justice system requires a prognostic
assessment of the risk of re-offending in a more or less broad sense,
numerous system places, in which the probability of committing new
crimes is potentially significant, can be identified:
 application, replacement and revocation of personal
precautionary measures (Article 274 et seq. of the Code of Criminal
Procedure);
 commutation (Article 133 of the Criminal Code);
 application of probation (Article 163 et seq. of the
Criminal Code);
 admission to release on temporary licence and

100
 alternative measures of detention (Articles 30-ter, 47 et seq. of Act
No 354 of 26th July 1975);
 application, replacement and revocation of personal
security measures (Article 203 et seq. of the Criminal Code);
 application, replacement and revocation of personal
prevention measures (Article 1 et seq. of Legislative Decree No
159 of 6th September 2011).
A prognostic assessment of the probability of re-offending can
be of importance, first of all, in the context of decisions on remand.
In the Italian system, as is well-known, the Code of Criminal
Procedure requires the existence of one of the three so-called pericula
libertatis listed in Article 274 of the Code of Criminal Procedure as
well as “serious evidence of guilt” (Article 273 of the Code of
Criminal Procedure) for the application of remand. The measures set
at the basis of the first two precautionary requirements – regarding the
danger of escape and danger for the acquisition and genuineness of the
evidence – are clearly connected with purposes of a trial nature. The
measure arranged for the third precautionary requirement (Article 274,
letter c, of the Code of Criminal Procedure) is on the other hand
directed towards dealing with the danger of “serious crimes”149 or
crimes “of the same type as the one being prosecuted” punished with
a maximum sentence by law of not less than four years (not less than
five years for the application of remand in prison) being committed,
thus responding to a function of special prevention that is not always
strictly connected with the proceedings in progress 150.

149
More precisely, article 274, letter c, of the Code of Criminal Procedure
refers to “serious crimes with the use of arms or other means of personal violence or
directed against the constitutional order, i.e. crimes of organised criminality”.
150
On the criticisms of the third precautionary requirement and its functional
connotation in the substantial sense, see F. CALLARI, Il periculum libertatis costituito
dal rischio di realizzazione di determinati reati e le misure cautelari: il fine giustifica
i mezzi?, in Dir. pen. cont., 12th November 2012, 1 et seq.

101
 Similar forms of pre-trial coercion of the investigated or
accused person, with overtones of a not strictly trial function, but
aimed instead at meeting the substantial requirements of special
prevention, are not however an Italian characteristic: measures with a
similar basis for their justification can in fact be found in all the main
European trial systems 151.
A second area of relevance for the probability of committing
new crimes must be identified in the time of proportioning the
sentence.
In the Italian system, the decision of the sentence to be applied
is given to the same judge as deliberated on the guilt of the accused.
The reference provision for proportioning the sentence (Article 133 of
the Criminal Code) lists a series of elements which the judge must take
into account, separated into profiles regarding the seriousness of the
crime and aspects regarding the guilty party’s so-called capacity to
offend. In particular, the so-called capacity to offend152 must be taken
from 1) the reasons for offending and the nature of the crime, 2) the
criminal and court records, 3) the conduct at the time of and after the
crime, 4) the person’s individual, family and social conditions of life.
The reference indicators for determining the sentence to apply
are not different, on a general level, to those provided for in the main
European systems: although a concept similar to the “capacity to
offend” provided by the Italian code cannot be found at a comparative
level, the French, Spanish and German criminal codes mention all the

151
See, for example, Article 144 of the French Code of Criminal Procedure, §
112a of the German Code of Criminal Procedure and Article 503 of the Spanish Ley
de Enjuiciamiento Criminal. On this issue, and also for the further regulatory and
bibliographical reference, see VOGLER, S. FOULADVAND, Standards for making factual
determinations in arrest and pre-trial detention: a comparative analysis of law and
practice, in Various Authors, Comparative Criminal Procedure, edited by J. E. Ross,
S.C. Thaman, Edward Elgar Publishing, 2016, 191.
152
On the notion of capacity to offend, see, for all, G. FIANDACA, E. MUSCO,
Diritto penale. Parte generale, VIII ed., Zanichelli, 2019, 796.

102
profiles concerning the personality and the socio-economic and family
situation of the author of the crime alongside aspects regarding the
gravity of the crime153.
The identity of the competent judge for the decision on the
guilt and for determining the sentence also seems to be a common trait
in the main continental systems, which are distinguished from
common law ones in which the judgement has a two-stage structure
and the sentence hearing – where the sentence to be applied is
determined – is carried out after the conviction, before a different
judge.
That said, within the decision on apportioning the crime, it is
at the time when what, in the Italian system, is called “capacity to
offend”, where the prognostic assessment of committing new crimes
appears to take on some significance, is weighed up. However, it is
easy to find a correspondence between the elements from which to
deduce the so-called capacity to offend and the factors normally
included in the input of the predictive algorithms used overseas.
A third context where the probability of recidivism comes into
prominence is the decision of the applicability of the special institutes,
for which the system requires, explicitly or implicitly, a negative
prognosis for the crime occurring again. Alongside probation (Articles
163 et seq. of the Criminal Code), for which this requirement is
expressly and clearly provided for, one can mention a whole set of
institutes that presuppose a lesser level of special prevention

153
For example, Article 132-1 of the French criminal code refers to the
personality of the author and his/her material, family and social situation (“la
personnalité de son auteur (…) sa situation matérielle, familiale et sociale”), § 46 of
the German criminal code includes the personal and economic conditions of the author
of the crime in the factors for proportioning the sentence (“seine persönlichen und
wirtschaftlichen Verhältnisse”), Article 66 ap. 6 of the Spanish criminal code provides
that personal circumstances of the person are taken into account (“las circunstancias
personales del delincuente”).

103
requirements: alternative sentences as of Act No. 689 of 24th October
1981 (custodial sentence of a minimum ten hours per day in prison,
probation, fine), the alternative measures to imprisonment provided
for by Act No. 354 of 26th July 1975 on the prison system (Article 47
et seq. of the Prison Regulations) and other prison benefits like
admission to release on temporary licence (Article 30-ter of the Prison
Regulations), and finally, conditional release (Article 176 of the
Criminal Code).
The comparative picture is shown to be rich in measures of a
similar tenor, responding to the same ratio and function and also,
therefore, based on a negative prognosis of new crimes being
committed154.
Further significant spheres for the probability of recidivism
can be found in the discipline of the safety measures 155 and the ante
delictum preventive measures 156 of a personal nature, for which a
judgement of “social danger”157 of the intended recipient of the
measure is required.
Institutes that come close to the personal safety measures in
Italian criminal law can be found in the German and Spanish
systems158, which also provide for a dual track of sentences and safety

154
See, for example, the similar institutes to probation provided for by the
French (Articles 132-29 and 132-40), German (§ 56) and Spanish (Article 80) criminal
codes.
155
For the personal safety measures, also in a comparative key, see for all M.
PELISSERO, Pericolosità sociale e doppio binario. Vecchi e nuovi modelli di
incapacitazione, Giappichelli, 2008, passim.
156
On social danger in the context of the discipline of personal preventive
measures, see lastly A. MARTINI, Essere pericolosi. Giudizi soggettivi e misure
personali, Giappichelli, 2017, spec. 81 et seq., 121 et seq.
157
For a recognition of the sphere of relevance for so-called social danger, see
lastly A. MARTINI, Essere pericolosi. Giudizi soggettivi e misure personali,
Giappichelli, 2017, passim.
158
See §§ 66 et seq. of the German Criminal Code and Article 96 et seq. of
the Spanish Criminal Code.

104
measures. Unlike the systems in France and the United Kingdom
which start from a monist setting that recognises the sentence as the
only criminal punishment and entrusts the health and administrative
system with taking charge of people released because they cannot be
charged and considered dangerous 159; the extra-penal nature of the
discipline does not at all exclude the existence of a component of
coercion, as, for example, particular procedural requirements are
provided for the release of patients receiving an order to be placed in
a psychiatric hospital160. Moreover, in the last decades, the tendency
to introduce safety measures for chargeable persons, which must
therefore be accumulated in the sentence, must also be recorded in
those systems, like the German161 and Spanish162 ones, which, unlike
the Italian one, were characterised by a so-called pure dual track, with
a clear distinction between intended recipients of the sentences (the
chargeable) and the safety measures (the non-chargeable). The areas
of the significance of so-called social danger thus appear ever more
extended.
With regard to the preventive measures governed by
Legislative Decree No. 159 of 6th September 2011, it is difficult to
find a corresponding institute in the comparative picture. On the other
hand, it concerns measures that, though already widely and sadly
known during Fascism and even before, have survived into the current
constitutional picture, being strictly tied to the experience of

159
M. PELISSERO, Pericolosità sociale e doppio binario, cit., spec. 151 et seq.
160
M. PELISSERO, Pericolosità sociale e doppio binario, cit., spec. 163 et seq.,
171 et seq.
161
M. PELISSERO, Pericolosità sociale e doppio binario, cit., spec. 161 et seq.,
255 et seq.
162
Various Authors, Manual de Introduccion al Derecho Penal, dir. J.
Lascuraín Sánchez, Agencía Estatal BOE, 2019, 280; X. ETXEBARRIA ZARRABEITIA,
Medidas de seguridad: presupuestos de su aplicación, in Cuadernos penales José
María Lindón, 10/2014, 125 et seq.

105
combating organised crime, particularly the Mafia.

2.2. Prediction of the decision-making result

Alongside the possibility of using tools that assess the

probability of a new crime being committed, the different potential of
the systems that aim to provide a prediction of the decision-making
result must be systematically collocated.
Although the main experience recorded at the moment is the
French one, which puts itself forward as a defence tool, it is not
difficult to imagine their potential for the work of the prosecution or
judges.
For example, the application of predictive justice tools for the
purposes of a sentence for a criminal action was theorised by the
Public Minister’s office in the Italian debate163.
In a system like the Italian one, which is imprinted to a canon
of the obligation to institute criminal proceedings 164, the choice
between drawing up the charges and requesting dismissal is based on
the legal parameter of the groundlessness of the notice of crime, which
is defined as unfitness of the elements acquired in the preliminary

163
C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale: molte
speranze e qualche equivoco, in Dir. pen. cont., 6/2019, 64.
164
A principle of the obligation to institute criminal proceedings is also
provided for in Germany (§ 152, II, StPO), but it is has been progressively watered
down by the introduction of margins of discretion for the Public Minister with regard
to minor crimes (Bagatelldelikte). On this subject, see Various Authors, Procedure
penali d’Europa, edited by M. Delmas-Marty, M. Chiavario, Cedam, 2001, 193. On
the other hand, for the margins of discretion that can concretely characterise the
activity of public ministers and the subject of the priority criteria in exercise criminal
action in Italy, see N. GALANTINI, Il principio di obbligatorietà dell’azione penale tra
interesse alla persecuzione penale e interesse all’efficienza giudiziaria, in Dir. pen.
cont., 23rd September 2019, 1 et seq.

106
investigations to support the prosecution in a court (Article 125 disp.
att. of the Code of Criminal Procedure). In this context, software
capable of indicating the probability of conviction at the outcome of
the debate in percentage terms could aid the Public Minister to assess,
precisely, the sustainability of the action being debated165.
In systems that provide for an opposite principle of
opportunity of criminal action, like the French one and the common
law ones166, a predictive tool like the one theorised about could be
inserted into the implementation of a programme for the activities of
the offices tasked with public prosecution: it could, in fact, serve to
direct the choices of the Public Minister in consideration of possible
criteria and guidelines which the offices – even in systems
characterised by discretion of action – can choose to equip themselves
with.
A further scenario for the use of predictive justice software by
a judge, which has been examined in the debate over the last few
years167, does not, for the moment, find significant confirmation on the
practical experience level in the criminal justice area. Software like
that developed in France for lawyers in the civil justice sector, capable
of indicating the probable result of the trial, could possibly be a tool to
support the judge in the customary activities of case-law research
which – even in systems that do not provide a binding nature of the

165
In any case, it must be stated that, even if a predictive tool capable of
offering an output for the probability of conviction in the presence of certain types of
evidential elements could be developed, the capacity of an algorithmic tool proceeding
to assess evidence or affirm the responsibility of a person “beyond all reasonable
doubt” must currently be excluded. Amplius, § 2.3.
166
Various Authors, Procedure penali d’Europa, cit., 119, 244.
167
See O. DI GIOVINE, Il judge-bot e le sequenze giuridiche in materia penale
(intelligenza artificiale e stabilizzazione giurisprudenziale), in Cass. pen., 3/2020, 951
et seq.; A. MASSARO, Determinatezza della norma penale e calcolabilità giuridica,
Editoriale Scientifica, 2020, 501; F. BASILE, Intelligenza artificiale e diritto penale,
cit., 14.

107
precedent – normally accompany the study stage preliminary to the
decision on the relevant legal points.

2.3. Further theoretical fields of application

A further field of application for computer systems worthy of

consideration is the calculation of the sentence (or the computation of
the prescriptive and trial terms) on the model of the Italian application
“Toga”.
It concerns a very cautious use of the potential of new
technologies, which would not in fact propose replacing the judge or
supporting him/her in the exercise of the discretion attributed by law:
the tool under examination would instead be limited to providing an
aid in the “arithmetical” concretisation of decisions already taken in
complete autonomy 168 or to check that he/she has not mistakenly
overstepped the legal limits of his/her discretion169.
A similar use of the technologies must therefore be framed as
a means of concretising the decisions on sentences and as a form of
(self)control of their legality, after the decision-making options
selected autonomously by the human judge.
Finally, a last potential sphere of relevance for AI in criminal
justice can be taken into consideration – if only to exclude it: the
assessment of the evidence and, more generally, the evidential
reasoning. In this regard, the judgement of the credibility of a witness

168
For example, the choice of applying the minimum (or maximum) sentence,
the recognition of an aggravation or mitigating circumstance, the judgement of
prevalence (or equivalence) of circumstances.
169
For example, if a sentence above the maximum by law is applied or
probation is granted outside the permitted limits, or if an increase for recidivism is
applied in cases not provided for.

108
or the assessment of gravity, precision and consistency of the evidence
are not at the moment easily compatible with the workings of an
algorithm (Article 192, paragraph 2, of the Code of Criminal
Procedure)170. Even the fundamental rule of “beyond all reasonable
doubt (BARD)” seems difficultly “calculable” by an algorithmic
tool171.

3. Criticisms and the risks of using AI technologies in the

criminal system

The sections above have attempted to reconstruct the picture

of the current – and some potential – uses of AI in criminal justice,
with some hints at systematic collocation in the context of the main
European systems. At this point, the critical issues arising in this
regard in the debate over the last few years can be examined.

3.1. Risk-assessment tools

As already mentioned, the systems for assessing the risk of

recidivism, already in use in the United States and also the subject of
experiments in Europe occupy a privileged place in the context of
reflection on the use of AI tools in criminal justice.
A lot of critical issues have come to light, on this side and the

170
In this sense, see F. BASILE, Intelligenza artificiale e diritto penale, cit., 15;
C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale, cit., 62.
171
In this sense, see F. BASILE, Intelligenza artificiale e diritto penale, cit., 16;
V. MANES, L’oracolo algoritmico e la giustizia penale: al bivio tra tecnologia e
tecnocrazia, in DisCrimen, 15th May 2020, 17. For the identification of some areas
for the use of algorithms on evidence, see J. NIEVA FENOLL, Intelligenza artificiale e
processo, cit., 70 et seq.

109
other side of the ocean, in the legal and other reflections.
Here, we propose to go through the highlighted problems,
putting them in relation to the general principles of law and the
criminal process, which would seem to be damaged by the
introduction of “intelligent” systems for assessing the probability of
recidivism. We will also offer a proposal for distinguishing between
critical issues connected with the IT/computation guise of the tools in
question and those linked to statistical/actuarial nature of the theories
at the basis of the “coded” algorithms in the predictive software172.

 De-individualisation of the decisions

One of the main problems that emerges from the use of risk-
assessment software is linked to the fact that the “predictive” output is
the product of an extension to the concrete case of the outcome of
statistical processing173 carried out on a sample of cases other than the
one in court. A decision taken (solely) on the response of the
algorithmic tool would therefore necessarily be the product of a
generalisation and would not derive from an individualised assessment
of the case in question.
If considered in the context of the stage of apportioning the
sentence, the criticism described could arrive at damaging principles
such as the guilt and the personality of criminal responsibility, which

172
Thus follows the interesting reflection of S. QUATTROCOLO, Equo processo
penale e sfide della società algoritmica, cit., 144.
173
See L. MALDONATO, Algoritmi predittivi e discrezionalità del giudice: una
nuova sfida per la giustizia penale, in Dir. pen. cont., 2/2019, 411; L. D’AGOSTINO,
Gli algoritmi predittivi per la commisurazione della pena. A proposito dell’esperienza
statunitense nel c.d. evidence-based sentencing, in Dir. pen. cont., 2/2019, 356; M.
GIALUZ, Quando la giustizia penale incontra l’intelligenza artificiale, cit., 21.

110
require an identification not only for ascertaining responsibility for the
crime but also for the sentencing.
If dropped into the context of the provisions for forms of
coercion in the precautionary stage (pre-trial), the risk of de-
individualising the decision could be in conflict with the guarantees
placed to protect personal freedom: the limitation of this right could in
fact only be ordered as extrema ratio, in the presence of certain
precautionary requirements, which must exist concretely, with
reference to the individual intended for the coercive measure.

 Discriminatory effects

The aforesaid risk of the de-individualisation of the decisions

and the use of statistical generalisation appears strictly linked to a
further criticism regarding the discriminatory effects 174 of the use of
risk assessment tools in criminal justice (probably the most well
known and discussed aspect, also at public opinion level).
If the algorithm is “trained” on a data set made up of historical
precedents, it could easily reproduce the discrimination implicit in the
conditions of greatest unease and social marginalization of certain
groups, even when the ethnic data - or, generally, the data for
belonging to a certain group - is not included in the information
initially entered into the system175. The operation of the most advanced
algorithmic tools is based – besides the data provided at programming
– on the connections identified by the same algorithm through
measuring the frequency of certain elements. It is therefore possible

174
See MALDONATO, Algoritmi predittivi e discrezionalità del giudice, cit.,
407; L. D’AGOSTINO, Gli algoritmi predittivi per la commisurazione della pena, cit.,
364; M. GIALUZ, Quando la giustizia penale incontra l’intelligenza artificiale, cit., 21.
175
L. MALDONATO, Algoritmi predittivi e discrezionalità del giudice, cit., 407.

111
that, by processing the data set at its disposal, the system recognises
connections between elements that have no connection on an
aetiological level176 and therefore produces discriminatory results.
A problem of compatibility with the principle of equality is
therefore posed, both with reference to the precautionary stage and
with regard to the time of apportioning the sentence.

 Author type criminal law and criminal determinism

Moreover, with regard to determining the punishment, one

cannot not notice that a predictive system conditioned by implicit
“prejudices” against specific categories of people may in substance
lead to an aggravation of the sentencing as a consequence of belonging
to a certain group: in other words, one would be punished not for what
one has done, but for what one is, according to a scheme that recalls
the model of “author type criminal law”177, contrary to the principles
of materiality and harm that characterise modern criminal systems
with a liberal matrix.
The criticisms just described – the de-individualisation of the
decision arising from the statistical matrix and the generalised nature
of the predictive result, the consequent discriminatory effects and the
tendency towards an author type criminal law – risk producing, on a
theoretical and cultural level, a slide towards a determinist approach,
which would re-propose the obsolete assumptions of the so-called
Positive School.

176
See X. RONSIN, V. LAMPOS, V. MAÎTREPIERRE, In-depth study on the use of
AI in judicial systems, notably AI applications processing judicial decisions and data,
cit., § 56 et seq., spec. § 66, 71.
177
See V. MANES, L’oracolo algoritmico e la giustizia penale, cit., 17; M.
GIALUZ, Quando la giustizia penale incontra l’intelligenza artificiale, cit., 21.

112
  Inaccessibility of the tool and opacity of its operation

Alongside the series of criticisms revolving around the

generalising and potentially discriminatory nature of the predictive
results, one of the most widely discussed problems concerning the use
of AI tools in the criminal system regards the poor transparency of the
mechanisms for operating the predictive algorithms.
The impossibility of the person affected by the judicial
decision to access the system is clearly in conflict with the principles
linked to the canon of fair trial and the prerogatives linked to the right
to defence: the proceedings would be carried out with a strong
limitation to the principles of cross-examination and equality of arms,
since the defence cannot dispute the validity of the predictive tool and
its result; the right to contest would then be substantially negated, with
the defence lacking the tools to take issue the reasoning passage that
refers to the output of the predictive software.
The inaccessibility of computational risk assessment tools is
normally traced back to the question of the commercial nature of many
of the predictive algorithms in use in the United States 178, the source
code of which is covered by trade secret and is therefore not disclosed
to either the judge or the parties (or any consultants that may intervene
in the proceedings). This aspect could perhaps be remedied by the
obligation to publish the source code and other relevant information
for the purposes of accessibility to the algorithm, through competitive
procedures for the providers of the predictive service or by generally

178
Supra, § 1.1. Some of the main algorithms used in the United States,
distinguishing between algorithms developed by public institutions, “commercial”
algorithms and non-profit algorithms, are examined in M. GIALUZ, Quando la giustizia
penale incontra l’intelligenza artificiale, cit., 5.

113
banning the use of “private” tools and reserving the development of
the software to be used in the criminal system to public
administrations.
However, some problems would remain.
In any case, a substantial negation of the principle of equality
of arms and the right to defence would be risked, against the high
specialisation required for the study of the algorithm to dispute its
validity and show its result to be false. The problem is not in fact new,
since similar criticisms characterise every proceeding where scientific
evidence is highlighted, which therefore requires the involvement of
specific professionalisms. Nevertheless, in the case of predictive
algorithms the problem could be made worse by the potential extent
of the area of application for these tools.
Moreover, it is not at all clear whether access to the source
code is enough to allow a complete knowledge of the tool’s operating
mechanisms, especially if it concerns algorithms capable of
reprocessing the data supplied at programming, developing forms of
“machine learning”179.

 Manipulability

A further criticism could also arise from the “computer

consistency” of predictive algorithms: one can imagine the eventuality
of undue alterations or intrusions into the system180 such as to
condition the decision-making result. By way of example, an undue

179
In this regard, “ontological inaccessibility of the operating mechanism” is
spoken about by L. MALDONATO, Algoritmi predittivi e discrezionalità del giudice,
cit., 408. See also J.P. DAVIS, Law without mind. AI, Ethics and Jurisprudence, in
University of San Francisco Law research Paper, 2018, 6.
180
This is remarked upon by O. DI GIOVINE, Il judge-bot e le sequenze
giuridiche in materia penale, cit., 957.

114
intervention on the software could be imagined: this might be aimed
at conditioning the judgement to be favourable to oneself or to be
unfavourable to a political or economic adversary or might even be
manipulation without a specific purpose, but which, in any case,
disrupts the decision-making process.
This would naturally concern a disease, for which suitable
protection measures could be prepared to guarantee the security of the
systems used in criminal justice. The harmful potential, particularly in
relation to the independence of judges, should however be mentioned.

 Cognitive biases and de-responsibilisation of the

judge

The entry of scientific-quantitative and automated risk

assessment tools in the criminal justice system could determine some
cognitive errors in the decision-making process of the judge, from
which, in the final analysis, an effect where the judge is de-
responsibilised could arise, even if the computer result is non-
binding181.
A first logical error, which a judge facing the result produced
by a predictive tool could run into, is linked to the cognitive bias
known as “anchoring”182, which consists of the influence of the first
information obtained on all the other subsequent passages in a
decision-making process.
Even surmising that this type of phenomenon could be
avoided or reduced, for example by only introducing the contribution

181
This risk is flagged by F. BASILE, Intelligenza artificiale e diritto penale,
cit., 22.
182
On this point, L. MALDONATO, Algoritmi predittivi e discrezionalità del
giudice, cit., 410; J. NIEVA FENOLL, Intelligenza artificiale e processo, cit., 38.

115
of the predictive tool after the independent decision of the judge 183,
there would still remain the risk of further cognitive error connected
to so-called automation bias184: when faced with computerised risk
assessment tools, there would be tendency to prefer the solutions
coming from the automated system over those reached without their
use.
Both the “cognitive prejudices” examined seem to lead to an
excessive and unjustified trust in the result produced by the predictive
tool. This means that, even in case of algorithms operating merely as
a consultation or an aid, this could in fact determine a substantial
delegation of the part of the decision-making process regarding the re-
offending risk assessment to an algorithmic tool. The latter effect
would seem to be one of the de-responsibilisation of the judge185.

 “Pollution” of the fact finding

A further criticism, also linked to the risk of cognitive errors,

concerns the possible pollution of the fact-finding judgement – and
hence the criminal responsibility of the person – by effect of the
acquisitions regarding the subjective conditions of the accused 186.

183
The use of predictive software as double check tool after the judge’s
assessment is speculated upon in V. MANES, L’oracolo algoritmico e la giustizia
penale, cit., 20.
184
On this point, G. UBERTIS, Intelligenza artificiale, giustizia penale,
controllo umano significativo, in Sist. pen., 11th November 2020, 4; P. COMOGLIO,
Prefazione, in J. NIEVA-FENOLL, Intelligenza artificiale e processo, Italian translation,
P. Comoglio, Giappichelli, 2019, X.
185
F. BASILE, Intelligenza artificiale e diritto penale, cit., 22.
186
In this regard “transfer of the judgement” is mentioned by O. DI GIOVINE ,
Il judge-bot e le sequenze giuridiche in materia penale, cit., 959. On this point, see
also L. MALDONATO, Algoritmi predittivi e discrezionalità del giudice, cit., 411; L.
D’AGOSTINO, Gli algoritmi predittivi per la commisurazione della pena, cit., 367; A.

116
 A similar contamination of the judgement of responsibility by
elements pertinent to the personality of the criminal could clash with
the paradigm of the criminal law of the fact and the principles of
materiality and harm, as well as the presumption of innocence, a
cardinal principle of the modern criminal process with a liberal stamp.
The possible transposition of assessments concerning the
personality of the author, in the context of fact-finding for his/her
responsibility for the fact being tried, is precisely what the legislators
of the Italian Code of Criminal Procedure intended to prevent with the
introduction of a provision like Article 220, paragraph 2, of the Code
of Criminal Procedure which forbids assessments aimed at
establishing the character or personality of the accused and, generally,
his/her “mental qualities independent of pathological causes”. In the
perspective of the legislator of 1988, a similar measure was made
necessary to prevent undue contamination and alterations of the
decision on the responsibility, in a system like the Italian one which,
like other European legislation, does not provide for separation of the
ascertainment of the fact and the sentencing187.

3.1.1. Some observations

In going over the criticisms that emerged from the debate on

risk assessment algorithms in the criminal field, it would seem that it
is possible to trace some problems back to their computational nature
and their “computer consistency” and others to the statistical-actuarial

GARAPON, J. LASSÈGUE, Justice digitale. Révolution graphique et rupture

anthropologique, Puf, 2018, 279.
187
For the two-stage structure that, on the other hand, characterises the
proceedings in common law systems and the autonomy of the sentencing stage, see L.
D’AGOSTINO, Gli algoritmi predittivi per la commisurazione della pena, cit., 367.

117
nature of the scientific theories put at the base of the algorithms which
are then “coded” in the software188.
So, for example, there is the use of generalisations based on
investigations of a statistical nature at the origin of the risk of de-
individualising the decision. The discriminatory results also seem to
go back, first and foremost, to the inevitable generalising effect
connected with the use of statistical laws 189.
Even the preoccupations of a broader range - regarding the
possible tendencies towards a criminal law of the author and
determinist drifts – seem to come from the use of generalising
methods, which lead to reasoning by categories or type of authors.
Otherwise, both the danger of predictive software being
manipulated and the risks for the right of defence arising from the
opacity of their operation seem to be directly connected to the
computerised nature of these systems.
As for the further critical points, connected with the risk of
cognitive biases, on one side, there is the so-called automation bias
which is determined precisely by the “computational nature” of the
tool and in the other cases (e.g. anchoring) it would seem to concern
cognitive errors that are in themselves independent of this element.
The observations above cannot and must not lead to a
minimisation of the problems of AI tools for the criminal justice
system: on the contrary, it concerns an attempt to make an in-depth
investigation into the origins of the criticisms reported, which can then
allow them to be contextualised within broader and well-known
problems – one for all, the entry of scientific evidence into the trial –

188
This concerns the setting suggested by S. QUATTROCOLO, Equo processo
penale e sfide della società algoritmica, cit., 144.
189
This however does not exclude that it is the computational capacity of the
tool allowing the database to be quickly processed, identifying the connections
intended to give shape to the statistical correlation between certain recurring elements
and the prognosis of recidivism.

118
and hence a better direction in the search for the most appropriate
precautions 190.

3.2. Tools for predicting the decision-making result

Some criticisms can also be highlighted with reference to the

use of tools predicting the decision-making result.
As regards the use that private parties could make of them, the
main problem to note is the one emerging from the debate in France:
in this particular context, the risk of undue checking of the judges was
highlighted, which may be susceptible of being translated into a breach
of the independence of the judiciary.
However, on closer inspection, the problem emerged
following the publication of a study that analysed the tendencies of the
competent individual judges on asylum law191 and not from the use of
predictive applications by law firms. In any case, the question is not
easily solved, considering the (equal) constitutional value of all the
interests in play: on the other hand, there are the guarantees regarding
knowability of jurisdictional provisions, which function under forms
of democratic control of the exercise of judicial power, as well as the
right to information and the freedom of expression of thought 192.
Because of this, the solution of the French legislator to omit the data
referring to the judge at the time the orders were published – and
criminalise the analysis and comparison such data193 – has not failed

190
Infra, § 5.
191
Supra, § 1.2.
192
On this point, see V. LAMPOS, V. MAÎTREPIERRE, In-depth study on the use
of AI in judicial systems, notably AI applications processing judicial decisions and
data, cit., § 43.
193
Supra, § 1.2.

119
to raise further misgivings194.
With regard to the use of decision-making predictive tools by
judges, the most perceived risk is that of generating a tendency to
conform with the dominant direction (so-called herd effect 195) which,
in the final analysis, leads to a standardisation of the interpretative
directions and decisions 196 and a de-responsibilisation of the judges 197.
On closer inspection, an effect of this kind happens, to a
certain extent, leaving aside the use of predictive tools of an
algorithmic nature: in the reasons for the sentencing, especially in
trials, there is in fact an ever wider significance of the references to
case-law precedents, even in systems like the Italian one, which do not
recognise the binding value of the precedent. This “decisional
conformism” may be partly the physiological effect of the
“nomophilachy” activities of the higher jurisdictions and partly a
pathological remedy to the need to make order in the complexity of
the legal points, at the same time managing a high workload.
In any case, the point to be stressed would seem not to be so
much the alignment to the dominant direction per se as, further up, the
methods of the software for identifying the “prevalent” opinion. In
fact, as already highlighted with reference to risk assessment
algorithms, it is possible that the system, working on the data set
supplied at programming, identifies connections between elements
that are actually unconnected in the legal reasoning that governs the
reasoning system for the orders examined.

194
M. LANGFORD, M.R. MADSEN, France Criminalises Research on Judges,
cit.
195
O. DI GIOVINE, Il judge-bot e le sequenze giuridiche in materia penale, cit.,
959.
196
The “ossification of the living law” is spoken about effectively by G.
UBERTIS, Intelligenza artificiale, giustizia penale, controllo umano significativo, cit.,
13.
197
F. BASILE, Intelligenza artificiale e diritto penale, cit., 22.

120
 Finally, the scenario can be considered – which has even been
put forward in doctrine – of algorithms being used by the Public
Minister as an aid in the choice between prosecution and dismissal 198.
In this regard, a real risk cannot seem to be recognised for the canon
of obligation to institute criminal proceedings – in systems like the
Italian one, where it is provided for – since the decision on dismissal
would in any case remain under competence of the “Giudice per le
Indagini Preliminari” [Preliminary Investigations Magistrate].

3.3. Other fields of application

No particular criticisms seem to arise for the use of computer

tools for calculating the sentence. As mentioned199, the system would
act after the decision-making options already chosen autonomously by
the human judge in exercising his/her discretion: it would be inserted
into the decision-making process only when every space for discretion
is exhausted, limiting itself to translating in arithmetical terms what
has been decided and allowing the judge a (self)check on the legal
limits to his/her power of discretion.

4. Legal limits on the entry of AI into the criminal system

The problematic aspects highlighted with regard to the entry

of AI systems into the criminal justice sector – as seen in the brief
analysis above – would seem to present some profiles that conflict
with the cardinal principles of criminal justice and the trial.
As regards the software and the risk assessment for

198
C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale, cit., 64.
199
Supra, § 2.3.

121
recidivism, criticisms have been made regarding the principle of the
personality of criminal responsibility, the principle of equality and the
principles of materiality and harm which give shape to the model of
criminal law of the fact; on a trial level, problems regarding the right
of defence and the canon of fair trial were examined, which would risk
being prejudiced by the lack of transparency of the software,
particularly that under private ownership and covered by trade secret.
As for the applications that offer a prediction of the decision-
making result, and in particular their use by judges to pose the problem
of a standardisation of decisions, which, in the absence of full clarity
regarding the operation of the algorithmic tool, would seem to be able
to corrode the principle of legality and the principle according to
which the judge is subject only to the law.
At this point, we need to examine the limits that currently exist
at an international level, in the context of single European and
domestic law, for the possible entry of AI into the criminal justice
system.
The “European Ethical Charter on the Use of Artificial
Intelligence in Judicial Systems and their Environment”, adopted by
the CEPEJ (European Commission for the Efficiency of Justice) of the
Council of Europe, undoubtedly has its significance on an
interpretative level.
In the context of European Union law, the regulations on the
protection of personal data take on particular importance in the subject
of the limits of AI in the criminal system.
Finally, with reference to the Italian system, the provision as
of Article 220, paragraph 2, of the Code of Criminal Procedure, which
could constitute a limit to the use of predictive software for the risk of
recidivism, must be examined.

122
 4.1. The Ethical Charter of the CEPEJ of the Council of
Europe

The “European Ethical Charter on the Use of Artificial

Intelligence in Judicial Systems and their Environment”200 was
adopted on 3rd December 2018 by the European Commission for the
Efficiency of Justice (CEPEJ) of the Council of Europe201.
The document affirms five principles:
1. the principle of respect for the fundamental rights
(ensure that the processing and implementation of the artificial
intelligence tools and services are compatible with the fundamental
rights);
2. the principle of non-discrimination (specifically
prevent the development and intensification of discrimination between
people and groups of people);
3. the principle of quality and security of the data (for the
processing of decisions and judicial data, use certified sources and
intangible data with models processed in multi-disciplinary modes, in
a technologically secure environment);
4. the principle of transparency, impartiality and fairness
(make the data processing methodologies accessible and
comprehensible, authorise external checks);
5. the principle of guarantee of human control
(precluding a prescriptive approach and ensuring that the users are
informed actors and have control over their choices).
Each of the five principles is accompanied by an explanatory

200
CEPEJ (2018)14
201
The content of the Charter, with a focus on the main questions for the
criminal sector, is analysed by S. QUATTROCOLO, Intelligenza artificiale e giustizia:
nella cornice della carta etica europea, gli spunti per un’urgente discussione tra
scienze penali e informatiche, in Leg. pen., 18th December 2018.

123
note that provides some clarifications.
With reference to the first principle, it refers to the necessary
respect for the fundamental rights provided for by the European
Convention on Human Rights (ECHR) for the protection of human
rights and fundamental freedoms and Convention No. 108 of the
Council of Europe on the protection against automated processing of
personal data. As regards the use of artificial intelligence tools
supporting the decision-making process of judges, specification is
made for the need to prevent breaches of the right of access to the
judge and the principles of fair trial (cross-examination and equality
of arms) and to respect the principles of legality and independence of
the judges. Finally, the preference for a human-rights-by-design
approach is expressed, according to which measures to prevent the
breach of the fundamental rights must be provided for directly from
the programming and learning stage of the tool
The explanatory note of the second principle specifies the
need to prevent the artificial intelligence tools from reproducing and
aggravating existing discriminations (with particular attention to those
regarding sensitive data, such as ethnic origin, social and economic
conditions, and political, religious and philosophical orientations)
leading to automatic and deterministic applications. Furthermore, it
also indicates the need to draw up measures to limit and, if possible,
neutralise the discriminatory effects and to disseminate adequate
knowledge of these risks amongst the people involved in the use of
artificial intelligence tools.
As regards the principle of quality and security of data, the
explanatory note recalls the need for the programmers of the artificial
intelligence tools for application in the justice sector to work in multi-
discipline groups, which allow collation with the legal professions and
researchers in the legal and social sciences field. It also makes
reference to the certification of provenance for the data entered into

124
the algorithms and automatic learning and their unmodifiability until
use by the software. Finally, it states the need that such systems are
kept and executed in secure environments which guarantee their
integrity and intangibility.
The explanatory note for the principle of transparency,
impartiality and fairness states the need to identify a balance between
intellectual property and the requirements of transparency,
impartiality and intellectual integrity for tools that can significantly
affect the life of people. In face of the difficulty of realising total
technical transparency because of the limit set by trade secret, the
possibility of publishing, in clear and comprehensible language, at
least some information on the nature of the services offered, the
operation of the tools and their margins of error was advanced;
secondly, a system of periodical certification by authorities or
independent experts for the artificial intelligence tools or preventive
consultancy was postulated.
Finally, with regard to the fifth and last principle, it was
specified, firstly, that the professionals in the justice system must be
able to review the judicial decisions and access the data used to obtain
a certain result and must not be bound by the output of the artificial
intelligence system, by reason of the specific characteristics of the
concrete case. Secondly, it was stated that there was a need for the
person concerned – i.e. the recipient of the decision – to be informed
in clear and comprehensible language, before or during the trial, of the
use of artificial intelligence tools, the binding or non-binding nature of
their output and his/her right to receive legal aid, to submit objections
and have access to a human judge.
The study published in first appendix of the Charter 202

202
X. RONSIN, V. LAMPOS, V. MAÎTREPIERRE, In-depth study on the use of AI
in judicial systems, notably AI applications processing judicial decisions and data,
cit.

125
dedicates special space to the mechanisms of how artificial
intelligence tools operate, highlighting that those currently available –
to be incorporated in the notion of weak or moderate AI – do not
reproduce the human reasoning model, but process statistical results
based on a very large quantity of data, thus presenting the risk of
identifying correlations that do not exist in reality 203. With specific
reference to the applications in criminal matters, besides bringing to
light the problems linked to the discriminatory effects 204, the study
highlights the importance of accessibility to the data in the algorithms
for the effectiveness of the right to defence and the guarantees of a fair
trial205. Moreover, special attention is reserved for the subject of the
protection of personal data206: in particular, the need for a prior study
on the risks of personal data processing, in order to prepare adequate
measures to prevent or minimise its impact on fundamental rights, is
highlighted; furthermore the so-called principle of openness in the data
processing, which prohibits the use of data for purposes other than the
initial ones, is affirmed.
The second appendix of the Charter divides the areas of use
for artificial intelligence in justice into four levels of caution: 1) uses
to be encouraged; 2) possible uses that require considerable
technological precautions; 3) uses to be considered following
additional scientific studies; 4) uses to be considered with the most
extreme reservations.
For what can interest the criminal justice sector, the use of
advanced case law research tools comes under the first group of
possible uses for artificial intelligence, which is considered free of
risks. On the other hand, the tools anticipating court decisions and the

203
Ibid., § 56 et seq., spec. § 66, 71.
204
Ibid., § 137.
205
Ibid., § 138.
206
Ibid., § 141 et seq.

126
processing of case law data aimed at profiling judges are included in
the third category of possible uses, to be considered following
additional scientific studies. Finally, the application of risk assessment
algorithms and the possible forms of locking into the mass of
precedents, based on a quantitative approach to case law, are the areas
for which the greatest level of caution is required.

4.2. European Union law: the limits to “decisions based solely

on automated processing”

In the context of European Union law, the discipline adopted

in 2016 on the protection of personal data (EU Regulation 679/2016
and Directive 2016/680) contains some provisions of great interest for
reflection on the possibility of introducing artificial intelligence
systems into the justice system: with reference to the criminal sector,
Article 11 of Directive 2016/680207 stands out; this concerns the
specific processing of personal data by “competent authorities for the
purposes of the prevention, investigation, detection or prosecution of
criminal offences or the execution of criminal penalties”.

207
Article 11 (“Automated individual decision-making “):
“1. Member States shall provide for a decision based solely on automated
processing, including profiling, which produces an adverse legal effect concerning
the data subject or significantly affects him or her, to be prohibited unless authorised
by Union or Member State law to which the controller is subject and which provides
appropriate safeguards for the rights and freedoms of the data subject, at least the
right to obtain human intervention on the part of the controller.
2. Decisions referred to in paragraph 1 of this Article shall not be based on
the special categories of personal data referred to in Article 10 unless suitable
measures to safeguard the data subject’s rights, freedoms and legitimate interests are
in place.
3. Profiling that results in discrimination against natural persons on the basis
of special categories of personal data referred to in Article 10 shall be prohibited, in
accordance with Union law”.

127
 The first paragraph of the provision in question obliges the
member states to provide for a prohibition on decisions based solely
on automated processing that produce negative legal effects or in any
case significantly affect the data subject 208. Such decisions can only be
admitted by provisions of Union or Member State law to which the
controller is subject and which provides appropriate safeguards for the
rights and freedoms of the data subject and among these, at least the
right to obtain human intervention on the part of the controller.
Moreover, the second paragraph prohibits decisions based on
the automated processing of data for criminal convictions and offences
or “connected security measures” (provided for by Article 10 of the
Regulation), which is only allowed when there are adequate measures
to protect the data subject.

208
A similar prohibition is provided for, in the context of more general
application, by Article 22 of EU Regulation 2016/679 (“Automated individual
decision-making, including profiling”):
“1. The data subject shall have the right not to be subject to a decision based
solely on automated processing, including profiling, which produces legal effects
concerning him or her or similarly significantly affects him or her.
2. Paragraph 1 shall not apply if the decision: (a) is necessary for entering
into, or performance of, a contract between the data subject and a data controller; (b)
is authorised by Union or Member State law to which the controller is subject and
which also lays down suitable measures to safeguard the data subject’s rights and
freedoms and legitimate interests; or c) is based on the data subject’s explicit consent.
3. In the cases referred to in points (a) and (c) of paragraph 2, the data
controller shall implement suitable measures to safeguard the data subject’s rights
and freedoms and legitimate interests, at least the right to obtain human intervention
on the part of the controller, to express his or her point of view and to contest the
decision.
4. Decisions referred to in paragraph 2 shall not be based on special
categories of personal data referred to in Article 9, paragraph 1, unless point (a) or
(g) of Article 9, paragraph 2, applies and suitable measures to safeguard the data
subject’s rights and freedoms and legitimate interests are in place”.
One can, however, recall that a prohibition of decisions based solely on
automated processing had already been provided for in Article 15 of Directive
95/46/EC.

128
 Finally, the third paragraph introduces an absolute prohibition
of profiling activities which produce discriminatory effects on natural
persons, based on the data as of Article 10 of the Regulation.
The interpretation of the notion of “decision based solely on
automated processing” is decisive for identifying the boundaries of the
scope of application of the prohibition as of Article 11 of the Directive
(as also Article 22 of the Regulation) 209.
A “minimum” reading is one which limits the subject of the
prohibition to the decisions entirely entrusted to a machine; the use of
the artificial intelligence tool would however be permitted as mere
support to the decision, provided that the intervention of the human
operator is not limited to simply signing the machine’s decision, but
consists of a substantial assessment of the decision210.
However, the regulatory formulation also allows a further

209
M. GIALUZ, Quando la giustizia penale incontra l’intelligenza artificiale,
cit., 16.
210
See M. BRKAN, Do Algorithms Rule the World? Algorithmic Decision-
Making in the Framework of the GDPR and Beyond, in Electronic Journal, 2017, 10:
“In order for the decision not to be based solely on automated processing, the human
judgment needs to be such as to verify the machine-generated decision and the human
should assess the substance of the decision and not be involved merely as another
(empty) procedural step. In other words, in order to escape the prohibition from
Article 22 GDPR or Article 11 of the Directive on Data Protection in Criminal
Matters, the human has to use the machine only as decision support, whereas the final
decision is taken by the human”.
In a similar sense, see the Linee guida sul processo decisionale automatizzato
relativo alle persone fisiche e sulla profilazione ai fini del regolamento 2016/679
(17/IT; WP 251 rev.01), 23, in the amended version adopted on 6th February 2018 by
the Article 29 data protection working party, set up by Article 29 of Directive
95/46/EC: “The controller cannot avoid the Article 22 provisions by fabricating
human involvement. For example, if someone routinely applies automatically
generated profiles to individuals without any actual influence on the result, this would
still be a decision based solely on automated processing. To qualify as human
involvement, the controller must ensure that any oversight of the decision is
meaningful, rather than just a token gesture. It should be carried out by someone who
has the authority and competence to change the decision. As part of the analysis, they
should consider all the relevant data”.

129
meaning to be attributed to the prohibition of decisions based on
automated processing: according to a suggestion from a part of the
doctrine, it would involve a rule to assess the evidence such that the
machine output is assigned a value of mere indication, always to be
corroborated with other elements of evidence211.

4.3. The limits of domestic law: the ban on criminological

expert opinions

In the Italian system, there would seem to be a significant limit

to the use of algorithmic tools for the assessment of the risk of re-
offending, like the ones widespread in the United States and under
experimentation in the United Kingdom.
It concerns a provision as of the second paragraph in Article
220 of the Code of Criminal Procedure, which prohibits: “expert
evidence to establish the habitualness or professionalism in the crime,
the tendency to commit crimes, the character and personality of the
accused and in general the psychological qualities independent of
pathological causes, unless provided for the purposes of executing the
sentence or safety measure”212.

211
M. GIALUZ, Quando la giustizia penale incontra l’intelligenza artificiale,
cit., 16; G. MALGIERI, G. COMANDÈ, Why a Right to Legibility of Automated Decision-
Making Exists in the General Data Protection Regulation, in International Data
Privacy Law, vol. 7, 1st November 2017, 14; V. MANES, L’oracolo algoritmico e la
giustizia penale, cit., 20.
212
On the prohibition of psychological and criminological expert evidence in
the Italian system, see F. ERAMO, Il divieto di perizie psicologiche nel processo
penale: una nuova conferma dalla Cassazione, in Dir. pen. proc., 7/2007, 931 et seq.;
P. MOSCARINI, La perizia psicologica e il ‘giusto processo’, in Dir. pen proc., 8/2006,
929 et seq.; P. MARTUCCI, Il contributo del criminologo nel processo penale: un
problema ancora aperto, in Dir. pen. proc., 6/2004, 744 et seq.; I. GIANNINI, Il
dibattito sull’ammissibilità della perizia e della consulenza criminologica nel
processo penale, in Rass. penit. crim., 3/2003, 87 et seq.; A. SAPONARO, L’esame della

130
 At the basis of the choice of the legislator of 1988, who
confirmed the direction of the procedure previously in force 213, there
were two fundamental preoccupations, corresponding to as many
guarantee requirements for the accused. First and foremost, the
prohibition of criminological and psychological expert evidence
seems to be inspired with the presumption of innocence, which could
be prejudiced by the expert evidence on the personality of the accused
who intervenes during the fact finding. Secondly, one can see the same
fear for the moral freedom of the accused that is at basis of Article 188
of the Code of Criminal Procedure in the background of the second
paragraph of Article 220 of the Code of Criminal Procedure214.
However, it must be noted that there are two provisions in the
Italian procedural system capable of being an opening valve for the
entry into the trial of fact-finding elements and judgement on the
personality of the accused. Article 194 of the Code of Criminal
Procedure, in prohibiting testimony on the morality of the accused,
admits it when “it concerns special facts, suitable for qualifying his/her
personality in relation to the crime and social danger”. Article 236 of

personalità del reo nel processo penale: evoluzione, involuzione, modelli alternativi,
prospettive, Cacucci, 2000, spec. 104 et seq.; P. RIVELLO, voce Perito e perizia, Dig.
disc. pen., IX, 474 et seq..; D. BIELLI, Periti e consulenti nel nuovo processo penale,
in Giust. pen., 2/1991, 65 et seq.; E. AMODIO, Perizia e consulenza tecnica nel quadro
probatorio del nuovo processo penale, in Cass. pen., 1989, 170 et seq.; G. TRANCHINA,
Il divieto di perizia psicologica sull’imputato: una limitazione sicuramente
anticostituzionale, in Riv. it. dir. proc. pen., 1971, 1325 et seq.
213
The formulation of Article 314, paragraph 2, of the Italian Code of
Criminal Procedure of 1930 was practically identical to that of today. “Expert
evidence to establish the habitualness or professionalism in the crime, the tendency to
commit crimes, the character and personality of the accused and in general the
psychological qualities independent of pathological causes are not permitted”.
214
Article 188 of the Italian Code of Criminal Procedure: “Methods, or
techniques that can influence the freedom of self-determination or alter the capacity
to recall and assess facts cannot be used, not even with the consent of the interested
person”.

131
the Code of Criminal Procedure permits the acquisition not just of
criminal record and irrevocable sentence certificates, but also
“documentation existing at the social services offices of public bodies
and supervision offices”, all documents that could contain
criminological or psychological opinions on the personality of the
subject. Case law, however, seems to be oriented towards limiting the
use of the documentation as of Article 236 only to judgements
regarding the recognition of so-called general mitigating
circumstances, sentencing, probation and non-mention215.
Moreover, a substantial exception to the prohibition of
criminological expert evidence in the judicial examination stage is
provided for by the discipline of the juvenile criminal trial216, where
“fact-finding on the juvenile” is allowed (Article 9 of Decree of the
President of the Republic No. 448 of 22nd September 1988 217), which
can also comprise the acquisition of expert opinions (paragraph 2).
Even if their function is mainly to allow alternatives to sentencing and
the trial to be assessed, one cannot exclude the use of the results of this
fact-finding for sentencing.
Finally, by virtue of the clause that opens the second
paragraph of Article 220 of the Code of Criminal Procedure, the use

215
Cass., sez. VI pen., 24th September 2013, no. 42823. On this point, S.
QUATTROCOLO, Questioni nuove e soluzioni antiche?, cit., 1763.
216
On this subject, C. DE LUCA, Gli accertamenti sulla personalità dell’autore
del reato minorenne e il divieto di perizia psicologica nel rito ordinario: riflessioni e
nuove prospettive, in Cass. pen., 2018, 2146 et seq.; L. CARCENI, Processo penale
minorile, in Enc. Dir., Agg., vol. IV, Giuffrè, 2000, 1018 et seq.
217
Article 9 of Decree of the President of the Republic 448/1988: “Fact-
finding on the personality of the juvenile. 1. The Public Minister and the judge acquire
elements regarding the personal, family, social and environmental conditions and
resources of the juvenile in order to ascertain whether he/she can be charged and the
degree of responsibility, assess the social significance as well as to arrange the
adequate penal measures and adopt any social provisions. 2. The Public Minister and
the judge can always use information from persons who have had relations with the
juvenile and hear the opinion of experts, without any formalities, for these purposes”.

132
of psychological and criminological assessments is of course allowed
in the executive stage. For that matter, the general formulation of the
law on the penal system (Act No. 354 of 26th July 1975) tends towards
a marked valorisation of the individualised treatment of the convict
and the internee based on the “scientific observation of the
personality”218.
It must be noted that a prohibition on criminological expert
evidence similar to the one in Article 220 of the Italian Code of
Criminal Procedure has not been found in the other main European
systems. Some possibilities for information and assessments on the
person’s personal situation to enter the trial are, on the other hand,
expressly provided for. For example, the French system provides for
“fast enquiries” (enquêtes sociales rapides) on the material, family and
social situation of the accused. These are aimed at checking the
practicability of determining penal options and identifying suitable
measures to promote social reintegration219. More in-depth
investigations on the personality of the accused and his/her material,

218
See Article 31, Act 354/1975 (“Individualisation of the treatment”). “The
penal treatment must meet the particular needs of each person’s personality. The
scientific observation of the convicted and internees is arranged to detect the physio-
psychological deficiencies and other causes of social maladjustment. The observation
is carried out at the beginning of the execution and continued during it. Based on the
results of the observations, indications regarding the re-education treatment to be
carried out are formulated for each convict person and internee and the related
programme is compiled. This is supplemented or modified according to the
requirements which are presented during the execution. The general and special
indications for the treatment are entered, together with the judicial, biographical and
health data, into the personal file in which the progress of the practical treatment and
its results are then noted. The collaboration of the convicts and internees with the
observation and treatment activities must be promoted”.
219
Various Authors, Procedure penali d’Europa, cit., 147. The “rapid
enquiries” can be arranged by the public prosecutor (Article 41, paragraph 8, of the
French Code of Criminal Procedure), the investigating magistrate (Article 81,
paragraph 7, of the French Code of Criminal Procedure) or the correctional court
vested with an application for immediate appearance (Article 396, paragraph 2, of the
French Code of Criminal Procedure).

133
family and social situation can then be arranged by the investigation
magistrate220; the results of this fact-finding go into a “personality
dossier, which has the function of providing the judicial authority with
elements on the conditions of the person’s past and present life and
“does not have the purpose of searching for evidence of his/her
guilt”221. The German system allows the Public Minister to request the
social services operators to hand over an “aid-to-the-decision” dossier
(Gerichtshilfe), which gathers together the elements regarding the
environment and social, family and psychological situation of the
accused222; this investigation is compulsory for persons between 18
and 21 years old223.
That said, it needs to be clarified whether the use of such tools
can be assimilated into the prohibited expert evidence to establish
whether the use of risk assessment software for recidivism comes
under Article 220 of the Italian Code of Criminal Procedure.
One part of the doctrine has expressed itself in favour of this
assimilation: according to this opinion, risk assessment tools like the
ones experimented with in the United States would represent a
technical-scientific assessment similar to expert evidence, since

220
Various Authors, Procedure penali d’Europa, cit., 147. See Article 81,
paragraph 6 of the French Code of Criminal Procedure.
221
See Article D16 of the French Code of Criminal Procedure. “L’enquête sur
la personnalité des personnes mises en examen ainsi que sur leur situation matérielle,
familiale ou sociale prévue à l’article 81, alinéa 6, du code de procédure pénale et les
examens, notamment médical et médico-psychologique, mentionnés à l’alinéa 7 du dit
article, constituent le dossier de personnalité de la personne mise en examen. Ce
dossier a pour objet de fournir à l’autorité judiciaire, sous une forme objective et sans
en tirer de conclusion touchant à l’affaire en cours, des éléments d’appréciation sur
le mode de vie passé et présent de la personne mise en examen. Il ne saurait avoir
pour but la recherche des preuves de la culpabilité”.
222
Various Authors, Procedure penali d’Europa, cit., 209. See § 160 of the
German Code of Criminal Procedure.
223
Various Authors, Procedure penali d’Europa, cit., 209. See § 38 of the
Jungengerichtsgesetz (JGG).

134
models and theories in psychology and criminology are used in the
programming stage for the algorithm224.

5. Final remarks. Spaces of application and precautions for

AI in criminal matters

The panorama of the algorithmic tools used in the context of

criminal justice and the further tendencies that have been recorded in
the civil sector have allowed us to identify two main uses of artificial
intelligence systems to which we pay give attention.
Firstly, the United States experience – which for the moment
finds correspondence limited to a single experiment in the United
Kingdom – has brought to light the use of tools aimed at assessing the
risk of re-offending, in the context of decisions based on the freedom
of people awaiting trial and those regarding punishment 225. Moreover,
it has also been seen how the prognosis of committing new crimes may
have a much wider area of significance extending to the application of
institutes such as probation, replacement penalties, alternative
measures to imprisonment and other prison benefits, as well as safety
and personal prevention measures 226.
However, the risk assessment software for re-offending is the
subject of wide discussion in doctrine and public debate because of the

224
In this sense, S. QUATTROCOLO, Questioni nuove e soluzioni antiche?, cit.,
1762. The same conclusion is reached by L. M ALDONATO, Algoritmi predittivi e
discrezionalità del giudice, cit., 412, who dwells on the inclusion of information
regarding the personality of the data subject, even taken from the replies given at
interrogation, in the software input. On the other hand, the similarity between the use
of risk assessment tools and prohibited expert criminological evidence is not at all
discounted by M. GIALUZ, Quando la giustizia penale incontra l’intelligenza
artificiale, cit., 20.
225
Supra, § 1.1.
226
Supra, § 2.1.

135
numerous criticisms that its use in the context of criminal justice seems
to present 227. Some of the problems brought to light can be traced back
to the question of the validity of the scientific theories and statistical
models at the basis of the operation of algorithmic tools: the
application to the concrete case of an output generated by the
processing of statistical correlations, according to a generalising
approach, could in fact lead to the de-individualisation of the judge
and, potentially, to discriminatory effects for minorities in conditions
of greatest discomfort228. On one hand, the computer and
computational nature of the tools in question is connected to the issue
of poor accessibility to the system and the codes which govern their
operation, with the risk of harming the right of defence and a fair trial,
especially in the case of algorithms that are private property and coved
by trade secret 229.
The second type of software emerging from the examination
of current experiences is that of predictive tools, developed mainly in
France, which propose offering a prediction of the decision-making
result230; although they are mainly used by law firms at the moment
and have been experimented with at very few judicial offices, these
applications nevertheless require a reflection on the potential and risks
of a wider use of them, in particular by judges 231.
The risk highlighted with reference to this type of algorithmic
systems concerns, on one hand, the independence of judges, in the case
of forms of processing and analysis of case law precedents that are
also aimed at the profiling of the single magistrate232. The wider

227
Supra, § 3.1.
228
Supra, § 3.1.1.
229
Supra, § 3.1.1.
230
Supra, § 1.2.
231
Supra, § 2.2.
232
Supra, § 3.2.

136
problem is however the one connected to the risk of standardising
decisions, as much for the effect of making the directions uniform as
the method of identifying the prevalent case law, which could be based
on merely quantitative data or statistical correlations erroneously
identified by the machine, starting from the processing of the elements
of previous cases 233.
Faced with the criticisms highlighted, the European law
picture, even soft law, seems to suggest an approach of extreme
caution with regard to the introduction of artificial intelligence tools
into the criminal justice system234. The five fundamental principles set
out by the CEPEJ Ethical Charter of the Council of Europe (respect of
fundamental rights, non-discrimination, quality and security of the
data, transparency, impartiality and fairness, guarantee of human
control) and the discipline of the European Union, which prohibits
decisions based solely on an automated processing of personal data,
offer an important starting point for the European Union or domestic
legislator who may intend to introduce regulations on the use of such
tools235. Moreover, as regards the predictive algorithms for the risk of
re-offending, the Italian trial system, unlike other systems, has always
excluded the possibility of executing criminological or psychological
expert evidence in the trial, by means of a provision like Article 220,
paragraph 2, of the Code of Criminal Procedure, which, according to
one part of the doctrine, could constitute an obstacle to the use of risk
assessment tools 236.
In light of the picture thus outlined, some final observations
can be formulated on possible precautions capable of combating the
multiple risks which were highlighted by the debate on the relationship

233
Supra, § 3.2.
234
Supra, § 4.1-4.2.
235
Supra, § 4.1-4.2.
236
Supra, § 4.3.

137
between artificial intelligence systems and criminal justice.
With reference to the instruments assessing the probability of
re-offending, the risk of the de-individualisation of the decision and
the potential discriminatory effects seem to require two fundamental
precautions: a serious multi-disciplinary debate on the quality of the
data and the validity of the theories and models generated by their
processing, as well as an adequate training of the judges called upon
to use the results of the algorithm. In particular, the legal operators,
who find themselves faced with the predictive output of an artificial
intelligence tool operating on the basis of statistical correlations,
should be warned of the generalising nature of this type of approach:
a clear awareness of the limits of the machine’s “reasoning” could in
fact avert the risk of automation biases and blind adhesion to the
“judgement” of the algorithmic system.
The transparency problem, which is tied to the need to
guarantee the effectiveness of the principles of fair trial and the
different prerogatives that give substance to the right of defence,
would seem to require, at least, clear incisive regulations in public law
which, as well as providing for stringent requirements on the selection
of service provider, also impose obligations of suitable disclosure to
allow the defence the widest possible access to the operating
mechanisms of privately owned systems: the preferable solution
would in any case be to reserve the development of the algorithms to
be used in court to public institutions.
On the other hand, the risk of the system being manipulated
must be met through suitable measures that guarantee the security of
the environment where the software in use in the criminal justice
system is kept and used.
Finally, in order to prevent possible prejudice to the
presumption of innocence, which could arise from the insertion of
assessment elements concerning the personality of the accused into the

138
trial, it could first of all be provided that such fact-finding data could
not be used for the purposes of establishing guilt. If this is not
sufficient to remove the risk of polluting the decision-making process
regarding responsibility for the deed, the scenario of introducing a
separation between the time of the decision of responsibility and the
time of sentencing, also in systems like the Italian one, could be
evaluated; it would however concern an intervention with a disruptive
impact at system level, thus to be considered with extreme caution
bearing in mind the overall equilibrium on which current trial
discipline is supported.
Moving on to consider artificial intelligence systems that
propose to offer a prediction of the decision-making result, the
fundamental precaution to be taken into consideration would seem to
be the training of the operators: judges, in particular, must be
adequately informed of how these types of tools operate so that they
can be aware, in their activities, of the fundamental difference between
human reasoning and the “reasoning” of the algorithm, which based
on correlations established on the basis of the statistical frequency of
certain groups of words and not on the weighing up of the arguments
at the basis of the previous case law precedents examined. As regards
the use of algorithms for the profiling of magistrates, the suggested
precaution, also suggested by the CEPEJ Ethical Charter, would seem
to corroborate the precautionary approach of the French legislator in
2019, which provided for the omission of data identifying the
individual judges and prohibited their use for the analysis and
comparison of their activities.
In any case, apart from the single precautions envisaged to
meet the criticisms that emerge from the application of artificial
intelligence tools in court, the fundamental need that must be recorded
is a wider awareness of the general mechanisms of how algorithms
work: only the operators’ greater awareness of the algorithmic process

139
and the consequent limits to its “product” may avert the risks of de-
individualisation of the decisions on the risk of re-offending and
“ossification” of case law. Moreover, further upstream, there seems to
be a need for a multi-disciplinary reflection capable of clarifying
whether the risk of false correlations and the problem of technical
inaccessibility to the operation of algorithms with “self-learning”
functions can be reduced to a measure considered acceptable for their
use in the context of criminal justice.

140
 CHAPTER IV

PREDICTIVE SOFTWARE AND CRIMINAL LAW

Lorenza Grossi

Table of Contents: 1. Introduction. – 2. What is artificial intelligence? – 3. A “system”

for criminal law. – 3.1 Predictive policing: Keycrime. – 3.2. Predictive
justice: COMPAS. – 4. Areas of criticism and future prospects.

1. Introduction

Alan Turing, the British logician and mathematician,

published an article called Computing machinery and intelligence in
1950237 – years before the date tentatively recorded as the “birth” of
artificial intelligence. In this article, known for having introduced what
would later pass into history as the “Turing Test”, Turing begins with
the provocative question Can machines think?238 and concludes with
the hope that the machine, one day in the future, would be capable of
competing with men in all purely intellectual fields239. Today, Turing’s

237
1956, the year when the conference from which it takes its name was held
at Dartmouth College, is usually identified as the date of birth of a new field of
research known, since then, as artificial intelligence, see J. MCCARTHY, M.L. MINSKY,
N. ROCHESTER, C.E. SHANNON, A proposal for the Dartmouth summer research
project on Artificial Intelligence, 31st August 1955, in www-formal.stanford.edu.
238
A.M. TURING, Computing machinery and intelligence, in Mind, LIX, 236,
10/1950, 433 et seq.
239
A.M. TURING, Computing machinery and intelligence, cit., 460: “We may
hope that machines will eventually compete with men in all purely intellectual fields.
But which are the best ones to start with? Even this is a difficult decision. Many people
think that a very abstract activity, like the playing of chess, would be best. It can also
be maintained that it is best to provide the machine with the best sense organs that

141
hope sounds like a prophecy: machines are capable of being
introduced into those sectors which, until a short time ago, were the
exclusive domain of man, promising results of the same quality, but in
considerably less time. Progress, in this particular historic moment,
cannot be imagined without the aid of these “machines”.
Even today we speak of “machines” and “thought” along the
lines of the coordinates identified by Turing: the two concepts have
evolved into the heading of “artificial intelligence” in our daily
vocabulary and, intuitively, the first problem is the nature of the
definition.

2. What is artificial intelligence?

According to a well-known definition, artificial intelligence is

that discipline which “studies the fundamental theories,
methodologies and techniques that allow the designing of hardware
systems and software program systems aimed at providing the
computer with performances, which, to the layman, would seem to the
exclusive domain of human intelligence”240. At this point, the

money can buy, and then teach it to understand and speak English. This process could
follow the normal teaching of a child. Things would be pointed out and named, etc.
Again I do not know what the right answer is, but I think both approaches should be
tried. We can only see a short distance ahead, but we can see plenty there that needs
to be done”.
240
F. AMIGONI, V. SCHIAFFONATI, M. SOMALVICO, in Enciclopedia della
Scienza e della Tecnica, 2008, in www.treccani.it, § 1.; see also M. CRAGLIA (Ed.), A.
ANNONI, P. BENCZUR, P.BERTOLDI, P.DELIPETREV, G. DE PRATO, C. FEIJOO, E.
FERNADEZ MACIAS, E. GOMEZ, M. IGLESIAS, H. JUNKLEWITZ, M. LÓPEZ COBO, B.
MARTENS, S. NASCIMENTO, S. NATIVI, A. POLVORA, I. SANCHEZ, S. TOLAN, I. TUOMI,
L. VESNIC ALUJEVIC, Artificial Intelligence – A European Perspective, EUR 29425
EN, Publications Office, 2018, according to whom: “AI is a generic term that refers
to any machine or algorithm that is capable of observing its environment, learning,
and based on the knowledge and experience gained, taking intelligent action or

142
argument’s centre of gravity will have to be shifted onto the definition
of “human intelligence”, in the absence of which one risks of falling
into a blind alley or, worse, a vicious circle. What the machine (or
rather, the system) wishes to emulate is not human intelligence per se,
but a type of (human) reasoning that can also be replicated by the
machine: inferential reasoning. A system can process a significant
quantity of data - and in relatively short time - placing itself as
adequate support for human activity, especially in a time, like this,
marked by an enormous bulk of recorded data (so-called Big Data, to
which the phenomenon of data mining, i.e. the capacity of a system to
extrapolate “useful” information from this indefinite jumble of data, is
strictly connected).
The reasoning through which the computer system processes
the data has a name: algorithm. This is nothing but a well-defined
procedure aimed at transforming input data into output data241.
In systems equipped with artificial intelligence, the machine
is required to be equipped with another algorithm – so-called meta-

proposing decisions. There are many different technologies that fall under this broad
AI definition. At the moment, ML techniques are the most widely used”; J.
MCCARTHY, M.L. MINSKY, N. ROCHESTER, C.E. SHANNON, A proposal for the
Dartmouth summer research project on Artificial Intelligence, cit., in the opinion of
whom: “The artificial intelligence problem is taken to be that of making a machine
behave in ways that would be called intelligent if a human were so behaving”. The
major definitions proposed and accepted internationally are shown by S. SAMOILI, M.
LÓPEZ COBO, E. GÓMEZ, G. DE PRATO, F. MARTÍNEZ-PLUMED, B. DELIPETREV, AI
Watch. Defining Artificial Intelligence. Towards an operational definition and
taxonomy of artificial intelligence, EUR 30117 EN, Publications Office of the
European Union, 2020, 17-18. On closer inspection, all the definitions, including the
one proposed here, suffer from extreme elasticity, so much so as to negate the very
work of definition. Despite this, it seems useful to construct a picture, even though
within such unstable boundaries, of the phenomenon in question, taking care to place
the accent on the operation and objectives that an artificial intelligence system
proposes.
241
T.H. CORMEN, C.E. LEISERSON, R.L. RIVEST, C. STEIN, Introduction to
Algorithms, 3rd ed., MIT Press, 2009, 5.

143
algorithm242 – capable of creating new algorithms from itself, which
are suitable for defining a new transformation process depending on
the problem represented from time to time. In fact, the machine is not
capable of posing questions, but it knows how to construct replies, also
through mechanisms unknown to its programmer. It is in this context
that the phenomenon of machine learning must be placed, through
which – not unlike human learning – the machine improves its
performances through experience; this also spreads its effects on the
predictability of the product, as it is plausible that the system operates
through procedures which are no longer controllable by its creator and
not always verifiable ex post243. It seems opportune to state that a
machine thus equipped allows output data to be obtained in conformity
with the problem represented and the input data supplied, acting
according to algorithms constructed by it: it is not capable of knowing

242
F. AMIGONI, V. SCHIAFFONATI, M. SOMALVICO, in Enciclopedia della
Scienza e della Tecnica, 2008, in www.treccani.it, § 3.
243
P. DOMINGOS, L’algoritmo definitivo: la macchina che impara da sola e il
futuro del nostro mondo, Bollati Boringhieri, 2016, pp. 7 et seq.; F. PASQUALE, The
black box society: The secret algorithms that control money and information, Harvard
University Press, 2015, 1 et seq.; emphasise this aspect, in doctrine: V. MANES,
L’oracolo algoritmico e la giustizia penale: al bivio tra tecnologia e tecnocrazia, in
Discrimen, 15th May 2020, 14-15, G. CONTISSA, G. LASAGNI, G. SARTOR, Quando a
decidere in materia penale sono (anche) algoritmi e IA: alla ricerca di un rimedio
effettivo, in Diritto di internet, 4/19, 620 e 630; R. BORSARI, Intelligenza Artificiale e
responsabilità penale: prime considerazioni, in MediaLaws, 3/2019, 264-265; J.
MILLAR, I. KERR, Delegation, relinquishment, and responsibility: The prospect of
expert robots, in R. Calo, A. M. Froomkin, I. Kerr (editor), Robot Law, Edward Elgar
Publishing, 2016, 107; C. CASONATO, Intelligenza artificiale e giustizia: potenzialità
e rischi, DPCE – online, 3/2020, 3379-3380; A. SIMONCINI, L’algoritmo
incostituzionale: intelligenza artificiale e il futuro delle libertà, BioLaw Journal –
Rivista di BioDiritto, 1/2019, 78-79; O. DI GIOVINE, Il judge-bot e le sequenze
giuridiche in materia penale (intelligenza artificiale e stabilizzazione
giurisprudenziale), in Cass. pen., 3/2020, 960-961.

144
or delving deeper into the problem, but is capable of solving it by
means of purely computational activity.
A humanisation of the machine that is thus valid for aiding the
human force in every context, promising greater levels of efficiency:
it is not by chance that this revolution was aimed at all the fields,
which, until just recently, were the exclusive domain of the human,
including criminal law.

3. A “system” for criminal law

An algorithm capable of predicting the risk of re-offending or

software that identifies when and where new crimes will be
committed: the dystopian future is already here. Reprising the
Turing’s question, machines, though in their own way, think – and do
it very well. This has generated a new and widespread use of
“intelligent” systems in the field of law in general and, for what is of
interest here, criminal law. This phenomenon, born and known for
some time across the Atlantic, has extended – almost – into all of
Europe, so much so as to bring about the issue, by the Council of
Europe, of the European Ethical Charter on the Use of Artificial
Intelligence in Judicial Systems and their Environment 244. The purpose
of the European Ethical Charter is to promote the “responsible” use of
artificial intelligence tools in the criminal trial, incorporating
principles that can define a model that the use, development and
regulation of these “intelligent algorithms” can be based on.

244
European Ethical Charter on the Use of Artificial Intelligence in Judicial
Systems and their Environment, adopted by the (CEPEJ) during its 31st plenary
meeting (Strasbourg, 3-4th December 2018), 11, in www.coe.org. On this point, for
all, S. QUATTROCOLO, Intelligenza Artificiale e giustizia: nella cornice della carta
etica europea, gli spunti per un’urgente discussione tra scienze penali e informatiche,
in La legislazione penale, 18th December 2018, 1 et seq.

145
 As a consequence, the current applications used in the
criminal trial are worthy, in this historic moment, of suitable
reflection: the objective of these is precisely the analysis of the
Keycrime, a predictive policing system used in and around Milan and
COMPAS, predictive justice software developed in 1998 and used in
the United States of America, in an attempt to highlight the advantages
and any areas of criticism that these systems present.
Artificial intelligence systems are usually identified under two
labels: predictive policing245, on one hand, and predictive justice 246, on
the other. The different nomenclature records the diversity of the
subject that the system is called upon to process. Under the first
profile, the predictive policing systems propose the objective of
predicting (and hence preventing) a crime being committed, through
the statistical processing of data 247, and/or identifying the author of a
series of already known crimes: their action is therefore limited to the
stage before prosecution, aiding the forces of law and order in the
investigative or prevention stage. Otherwise, as regards predictive
justice, the different stages of the proceedings and the criminal trial
are the subject-matter of the software: as we will be able to specify
below, these applications have been used, particularly, to predict the
risk of the crime being repeated - in the dual context of pretrial release
and sentencing.
Both tools, obviously, have the capacity to affect the decisions
of the judge: the first, directly, by entering into the trial under the form
of evidence (in observance of the instructions in force on the point) 248;

245
Supra, A. GIRALDI, Chapter II of this research.
246
Supra, L. NOTARO, Chapter III of this research.
247
For an in-depth examination of the criminological theories at the basis of
similar tools, see R. PELLICCIA, Polizia Predittiva: il futuro della prevenzione
criminale?, in Cyberlaws, 9th May 2019.
248
C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale, in Dir.
pen. cont. – Riv. trim., 6/2019, 58.

146
the second indirectly - since it has the capacity to directly condition
the years or length of the sentence to be applied in the concrete case.
One last clarification: the adjective common to the two
phenomena is predictive and, as one knows, the future cannot be
predicted in certain terms, but only highly or not highly probable ones.
Currently, what makes the software prediction certain is that man acts
according to its expectations.

3.1 Predictive policing: Keycrime

The use of artificial intelligence software has been shown to

be particularly effective in investigative and/or prevention activities,
where there is a very large number of recorded data abstractly useful
for “solving” a case and the ability to cross check the information is
becoming increasingly less nimble for a human being 249. At the
moment, there are many types of software used in this field; by way
of example, the following should be noted: Situation-Aware Public
security Evaluation platform (in China), Crime Anticipation System
(Amsterdam), PredPol (United Kingdom and many states of
America), PreCobs (Germany and Switzerland) 250, Keycrime (Italy).
Considering the excellent results obtained, both on a
preventive and an investigative front, it is reasonable to presume that,
in a little while, every individual state will equip itself with its own
system. Nevertheless, it should be stated right from the start that the

249
In this sense, see A. MASSARO, Determinatezza della norma penale e
calcolabilità giuridica, Editoriale scientifica, 2020, 493-494; F. BASILE, Intelligenza
artificiale e diritto penale: quattro possibili percorsi di indagine, in Diritto penale e
uomo, 10/2019, 10.
250
S. HANNEM, C. B. SANDERS, C. J. SCHNEIDER, A. DOYLE, T. CHRISTENSEN ,
Security and Risk Technologies in Criminal Justice: Critical Perspectives, Canadian
Scholars, 2019.

147
majority of these tools are developed and owned by private companies,
which for reasons covered by intellectual property rights 251, omit to
provide the individual authorities with the software source code – but
more of this, later.
The predictive policing systems can be classified into two
categories depending on the premises on which their purposes are
built:
- systems aimed at identifying so-called hot
spots, i.e. places where, according to statistical calculations, it
is highly probable that new crimes may be committed; a
prediction, which, in that sense, allows the controls on “high
risk” areas to be intensified252;
- systems aimed at identifying the authors of
serial crimes: in these cases, the system permits appreciable
results in the cases where the crime committed is part of a

251
On the problems raised by the absence of transparency, see infra § 4.
252
On the misgivings raised by the so-called criminal mapping tools, see X.
RONSIN, V. LAMPOS, A. MAIT̂ REPIERRE, Questioni specifiche della giustizia penale:
prevenzione del reato, del rischio di recidiva e valutazione del livello di pericolosità,
Appendice I: Studio approfondito sull’utilizzo dell’intelligenza artificiale [IA] nei
sistemi giudiziari, segnatamente delle applicazioni dell’intelligenza artificiale al
trattamento di decisioni e dati giudiziari, in www.coe.int, 35: “the predictive
capabilities of these tools, which show their limitations with regard to crimes of a less
regular nature or targeting different locations, such as terrorism, must be put into
perspective. In addition, one of their weaknesses is the effect of “vicious circles” and
“self-fulfilling prophecies”: neighbourhoods considered at risk attract more police
attention and police detect more crime, which leads to excessive police surveillance
of the communities living in them”. On this point, the criticisms raised with reference
to Predpol, a system widely used in the Unites States and Great Britain, are
particularly significant. This was labelled as discriminatory, as some areas inhabited
by people of colour were identified as hot spots, even when the crimes, on a more
careful analysis, were recorded over a much more extensive area. In this sense, see K.
LUM, W. ISAAC, To predict and serve?, in Significance, vol. 13, 10/2016, 14-19; K.
LUM, Predictive Policing Reinforces Police Bias, in Human Rights Data Analysis
Group, available at www.hrdag.org; A. MASSARO, Determinatezza della norma
penale e calcolabilità giuridica, Editoriale scientifica, 2020, 494.

148
 series (crime linking); once the data in the chain is processed,
it is possible to predict where and when the next “job” will be
committed, prevent it from being committed and attribute the
whole series to its author.
The Keycrime software, which has been used, since 2008, in
the Municipality of Milan and, since 2009, over the whole district,
belongs to the second category. The history of Keycrime begins in
2004. Mario Venturi is assistant chief at the Milan Police
Headquarters: during his work, he found himself analysing an
enormous amount of data and, on careful analysis, found how some
crimes, far from representing a monad in the life of their author, place
themselves as a ring in a longer chain. From there, the intuition: if all
these crimes were properly examined and cross checked against each
other, the possibility of reaching a solution to the case, at the very least
in identifying serial authors, would certainly be greater 253. Keycrime
therefore works with exclusive reference to serial behaviours and, in
this context, points in two directions, opposite and complementary at
the same time: on one hand, the purpose of the application is to retrace
the fil rouge that connects and binds the series and, on the other, once
the data characterising the modes of conduct and its author have been
acquired and processed, it is capable of predicting when and where the
next crime will be committed254.

253
IBM in#300sec #ThinkMilano 2018 with Mario Venturi, KeyCrime,
available at www.keycrime.com; M. VENTURI, KeyCrime© – La chiave del crimine,
in PrimoPiano, 12/2014, 4, available at www.onap-profiling.org; C. MORABITO, La
chiave del crimine, in www.poliziadistato.it, July 2015, 36-38;
254
M. VENTURI, KeyCrime© – La chiave del crimine, in PrimoPiano,
12/2014, 4, available at www.onap-profiling.org; V. MANES, L’oracolo algoritmico e
la giustizia penale: al bivio tra tecnologia e tecnocrazia, cit., 6-7; F. BASILE,
Intelligenza artificiale e diritto penale: quattro possibili percorsi di indagine, cit., 12-
13; C. PARODI, V. SELLAROLI, Sistema penale e intelligenza artificiale, cit., 56; G.
MASTROBUONI, Crime is Terribly Revealing: Information Technology and Police
Productivity, in The Review of Economic Studies, vol. 87, 11/ 2020, 2732.

149
 Keycrime, since the time of its development, has always been
used to identify the (serial) authors of robberies against trading
enterprises and banks – in fact, it has been shown how 70% of this
type of robbery can be traced back to serial conduct 255. Considering
the excellent results obtained, from two years ago to now, experiments
have also been made with the application for burglaries in flats 256;
however, on closer inspection and given the operation of the software,
excellent results could be obtained in all those cases where the
repeating of the conduct is a constituent element of the crime or in all
those crimes which, statistically, are shown with a certain “seriality”
– think, just to give an example, of child pornography crimes or, more
generally, sexual crimes.
At the start of the experiments, the application was capable of
storing and processing up to 11,000 input data257; now, under the guise
implemented by the name delia, the calculating capacities have
increased considerably, managing to cross check up to 1.5 million
data258. Pausing on this aspect, the algorithm is formed by input data
regarding:
- the physical characteristics of the author:
build, colour of hair, presumed age, sex, ethnicity and accent;

255
G. MASTROBUONI, Crime is Terribly Revealing: Information Technology
and Police Productivity, cit., 2753; C. SABELLI, Scacco alla malavita: arriva
l’algoritmo che prevede i reati, in Il Messaggero, 11th December 2017.
256
Public Prosecutor at the Court of Appeal of Milan-Public Prosecutor at the
Court of Milan, Bilancio di responsabilità sociale 2018, in
www.ca.milano.giustizia.it;
257
M. VENTURI, KeyCrime© – La chiave del crimine, cit. 5; G. MASTROBUONI ,
Crime is Terribly Revealing: Information Technology and Police Productivity, cit.,
2741; C. SCOVACRICCHI, Quando il poliziotto diventa startupper. The History of
Keycrime in www.startupmagazine.it, 11th April 2018.
258
Data available at www.keycrime.com.

150
 - all the circumstances in which the conduct
was practised: use of firearm, type of business robbed, method
of escape, vehicle and number plate259.
These data, not unlike what happens during the preliminary
investigations, are acquired by the judicial police at the time of brief
eye witness information or through the acquisition of images and/or
videos from the surveillance systems: what changes, compared to the
“orthodox” method, concerns the stage immediately after, when the
data are transferred into the software which searches for elements of
contact with the other data in the dataset.
Using the input data, the application is capable of providing as
output: (i) the connection between crimes, identifying the series that
can be attributed to the same author; (ii) the prediction for the latter’s
“next job” (when, where and how the next crime should occur)260. This
is possible, in the first stage, though inductive analysis, aimed at
comparing and analysing the data in the software in order to identify
the common elements that define the series 261; the second stage, on the
other hand, concentrates on a deductive analysis, which, through
observation of the data regarding the concrete case, is capable of
identifying the leitmotiv which characterises the author’s modus

259
G. MASTROBUONI, Crime is Terribly Revealing: Information Technology
and Police Productivity, cit., 2741; ID. (Lead Academic) Impact: Imagine being able
to predict a crime in the future, Research case study by University of Essex, in
www.essex.ac.uk; C. MORABITO, La chiave del crimine, cit., July 2015, 36-38.
260
C. MORABITO, La chiave del crimine, cit., July 2015, 36-38; R. PELLICCIA,
Polizia Predittiva: il futuro della prevenzione criminale?, cit.: “The study (by the
University of Essex) has highlighted what appears a consistent strong point of
Keycrime compared to its competitors. Whilst these work on a purely statistical basis,
indicating where, when and what type of crime will be committed, Keycrime also
offers to define the how, thanks to an analysis of the author’s modus operandi and
his/her psychological traits”.
261
M. VENTURI, KeyCrime© – La chiave del crimine, cit. 5-6;

151
operandi and predicts where, when and how the new crime will take
place262.
In fact, Keycrime implements a multi-discipline analysis,
operating transversally via mechanisms that belong to different sectors
such as mathematics, statistics, behavioural psychology and geospatial
analysis263.
The data shown below264, which were obtained from the
Municipality of Milan, in the years 2008-2017, and the District of
Milan, in the years 2009-2017, clearly show the effectiveness of the
application.

262
M. VENTURI, KeyCrime© – La chiave del crimine, cit. 6;
263
M. VENTURI, KeyCrime© – La chiave del crimine, cit. 5;
264
Available at www.keycrime.com.

152
153
 The graphs show how the application achieved considerable
results on two distinct levels: (i) firstly, one sees a clear improvement
between the possibilities of solving the case pre Keycrime and post
Keycrime (from 10% to 63% for the area in the Municipality of Milan
and 25% to 60% for the district); (ii) secondly, one can see how,
through use, it is the application itself that improves its performance,
with an increase in the cases solved of 16% and 22%, for the
Municipality and the District of Milan respectively, with the logical
consequence of a decrease in the same type of crime being committed
in those areas (from 28% to 58% for the Municipality of Milan and
from 25% to 89% in the district area).
At least two different factors contribute to the Keycrime’s
encouraging results: first and foremost, one improvement in this sense
is due to machine learning mechanisms 265, so that the machine learns
from experience; the other is that the mechanism of Keycrime, based
on the identification of a series of crimes, behaves in such a way that
the more data entered the better the performance of the software will
be in identifying the series and predicting the author’s behaviour. The
longer the series, the greater the possibilities of predicting the future
behaviour are266: which, clearly, has dissuasive effects on new crimes
being committed, leading moreover to a saving calculated at 2.5
million euros a year for civil society267.
Despite the results, which are certainly surprising, it is best not
let oneself be carried away by easy enthusiasm. Putting off the analysis

265
Supra § 2
266
G. MASTROBUONI, Crime is Terribly Revealing: Information Technology
and Police Productivity, cit., 2753; C. SABELLI, Scacco alla malavita: arriva
l’algoritmo che prevede i reati, in Il Messaggero, 11th December 2017.
267
G. MASTROBUONI, Crime is Terribly Revealing: Information Technology
and Police Productivity, cit., 2757; Predictive analytics, Oltre Venture investe in
KeyCrime, available at www.keycrime.com.

154
of (possible) criticisms that software like this can leave until later 268,
it is appropriate, at this time, to stress how these systems are not
exempt from criticism, though to a considerably lesser extent than the
misgivings raised by predictive justice applications.

3.2 Predictive justice: COMPAS

The analyses carried out so far start from an implicit premise:

man, though free, is predictable. The behaviour of the robber is
predictable, as is the reasoning of the judge, who explains to be in
compliance with binding “rules of the game”, which drive him/her
forward in the ascertainment of the accused’s criminal
responsibility269.
The use of predictive justice software is inserted in a more
widespread tendency, the subject of appropriate analysis by
criminologists and psychologists and aimed at outlining the criminal
personality of the criminal so that a prognostic judgement can be
issued on the risk of re-offending270. Thus, it is not surprising how the
applications developed – and used mainly in common law systems –
are moving in the direction of predicting the risk of the crime being
repeated by the person under investigation/the accused, on the dual
ground of pretrial release and sentencing.
Under this profile, it appears opportune to make a distinction
right from the start: one consideration is to state the risk of re-
offending for a person whose criminal responsibility has been declared

268
Infra § 4.
269
For an in-depth study of how judges decide and the predictability of the
product of their reasoning, see J. NIEVA-FENOLL, Intelligenza artificiale e processo,
translated by P. Comoglio, Giappichelli, 2019, 31 et seq.
270
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 57, to which you
are referred for appropriate bibliographical references.

155
with the decision therefore being based on the elements of proof
arising during the trial; on the other hand, the same operation is
completely different with reference to a person who is presumed
innocent until proven otherwise, leaving wide margins of decision to
software, the operation of which, as already discussed, cannot be
known271.
Among the many tools drawn up in this context, the sector
leader is certainly COMPAS (Correctional Offender Management
Profiling for Alternative Sanctions), developed in 1998 by the private
company Northpointe Inc. (now Equivant)272.
The application proposes, on statistical bases and in the
manners best gone into depth below, the risk of the person under
investigation/the accused re-offending under three distinct aspects: (i)
pretrial risk; (ii) general recidivism; (iii) violent recidivism. Alongside
the strictly “predictive” function, the system also draws up a “need
scale”, suitable for defining the author’s profile and his/her
rehabilitation requirements 273.

271
The reference is again to the impossibility of accessing the source code of
the software for reasons of intellectual property rights. For the different application of
the systems at the pretrial release and sentencing stage, see J. NIEVA-FENOLL,
Intelligenza artificiale e processo, cit., 57, who does not hesitate to state how: “In the
first case, one starts from a premise that has greater possibility of success, given that
the results will also be supported by the trial result which concluded with a conviction.
In the second case, however, the uncertainty is maximum”.
272
Practitioner’s Guide to COMPAS Core, 4th April 2019, 1, available on
www.equivant.com.
273
Practitioner’s Guide to COMPAS Core, cit., 21, available on
www.equivant.com: “The need scales are not meant to be predictive but aim simply
and accurately to describe the offender along dimensions relevant for correctional
practice. Research findings indicate that individuals involved in the criminal justice
system often have problems and deficits in the domains of education, housing,
employment, substance abuse, relationships, and cognition. The need scales should be
valid and reliable measures of constructs in these domains and other aspects of the
person-in-environment that represent potential targets for interventions. The need

156
 More particularly, the input data are formed by information
given by the person under investigation/the accused, through a
questionnaire filled in by him/her, and by extrapolating personal
details and other characteristics directly from the dossier. By
comparing these data with those of similar profiles already in the
software dataset, the algorithm processes the following as output: (i) a
decimal number 274, which corresponds to the “degree” of risk of the
crime being committed for all three of categories taken into
examination by the system (pretrial risk scale, general recidivism
scale, violent recidivism scale); the need scales. The questionnaire
given to the person under investigation/the accused275 consists of 137
items broken down into the following themes:
- pending decisions;
- criminal history;
- non-compliance;
- family criminality;
- company frequented;
- substance abuse;
- residence/stability;
- level of education;
- work experiences and financial situation;

scales guide individualized decisions for case planning, including identifying targets
and choosing interventions”.
274
See Practitioner’s Guide to COMPAS Core, cit., 8: “In general the decile
rank has the following interpretation:
 1 – 4: scale score is low relative to other offenders in norm group.
 5 – 7: scale score is medium relative to other offenders in norm
group.
 8 – 10: scale score is high relative to other offenders in norm
group”.
275
Available at the following link:
https://www.documentcloud.org/documents/2702103-Sample-Risk-Assessment-
COMPAS-CORE.html

157
 - leisure/recreation;
- social isolation;
- criminal personality;
- anger management;
- criminal attitudes.
The profiling, no need to say, is very accurate so much so as
to glimpse the actual logic of a (never) overshadowed positive school
in the precise questions aimed at defining the author’s history and
social context. The data, found this way, contribute to determining the
decimal number for the risk of recidivism; nevertheless, the process
through which this operation unfolds is unknown: the company has
omitted to reveal its content for reasons covered by intellectual
property276. The misgivings caused by this lack of transparency
exploded in the celebrated case of State v. Loomis277, from which it is
fitting to take the steps to highlight, though in a dynamic perspective,

276
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 57; G. CONTISSA,
G. LASAGNI, G. SARTOR, Quando a decidere in materia penale sono (anche) algoritmi
e IA: alla ricerca di un rimedio effettivo, cit, 621; D. KEHL, P. GUO, S. KESSLER,
Algorithms in the Criminal Justice System: Assessing the Use of Risk Assessments in
Sentencing (Responsive Communities Initiative, Berkman Klein Center for Internet &
Society, Harvard Law School), 7/2017, 11, available at www.dash.harvard.edu.
277
State v. Loomis, 881 N.W.2d 749 (Wisc. 2016). For an in-depth
examination, see D. KEHL, P. GUO, S. KESSLER, Algorithms in the Criminal Justice
System: Assessing the Use of Risk Assessments in Sentencing, cit., 18; E. CHIEL, Secret
Algorithms That Predict Future Criminals Get a Thumbs Up From Wisconsin
Supreme Court, in www.splinternews.com, 27th July 2016; J. NIEVA-FENOLL,
Intelligenza artificiale e processo, cit., 61-62; K. FREEMAN, Algorithmic Injustice:
How the Wisconsin Supreme Court Failed to Protect Due Process Rights in State v.
Loomis, in N.C. J.L. & Tech., 12/2016 75 et seq., available in
www.scholarship.law.unc.edu. In the Italian doctrine, see A. MASSARO,
Determinatezza della norma penale e calcolabilità giuridica, Editoriale scientifica,
2020, 497; S. QUATTROCOLO, Quesiti nuovi e soluzioni antiche?, in Cass. pen.,
4/2019, 1750-1753; A. SIMONCINI, L’algoritmo incostituzionale: intelligenza
artificiale e il futuro delle libertà, cit., 78-79; S. CARRER, Se l’amicus curiae è un
algoritmo: il chiacchierato caso Loomis alla Corte Suprema del Wisconsin, in
Giurisprudenza Penale Web, 24th April 2019.

158
the major criticisms that are hidden behind the indiscriminate (and, for
reasons better shown later, discriminatory) use of these applications.
Eric Loomis was an American citizen, who, in 2013, was
stopped whilst driving a car involved in a shooting. There were five
charges against him, Loomis pleaded guilty to the two less serious
crimes. In order to determine the sentence, the court asked for the
Presentence Investigation Report (PSIR)278. The PSIR included the
results obtained through COMPAS, which showed a very high risk of
recidivism in all three sectors assessed by the application (pretrial risk
scale, general recidivism scale, violent recidivism scale). Following
this, the court issued a particularly severe sentence for the offences he
was charged with: six years in prison and five years’ probation.
Loomis, after seeing his motion for post-conviction release rejected279,
disputed the decision before the Supreme Court of Wisconsin. In this
hearing, the accused complained of the breach of his constitutional
right to a fair trial, identifying three reasons for grievance: (i) breach
of his right to be convicted on the basis of accurate information (since
the application was protected by trade secret, Loomis was prevented
from assessing the accuracy of the score); (ii) breach of his right to an
individual sentence (COMPAS, as stated above, processes the data in
purely statistical terms, comparing them with similar profiles in the
dataset); (iii) the tool, in defining the risk of recidivism, had used the
gender data in a contentious manner – estimating greater risks for men

278
The Presentence Investigation Report (PSIR) is a report prepared, at the
request of the court, by a probation officer, aimed at discovering the criminal, family
and social history (including the educational background) of a person convicted for a
crime. It summarises the basic information the court needs for determining the
appropriate sentence. The sentence of a convicted person is increased or reduced after
the PSIR has been examined.
279
See S. QUATTROCOLO, Quesiti nuovi e soluzioni antiche?, cit., 1751-1752.

159
than women280. The Supreme Court of Wisconsin pronounced with
notes that, in a certain way, reprise those used by the European Court
of Human Rights in the well-known case of Murray v United
Kingdom281: in fact, by rejecting the certiorari request, the non-
exclusive use of COMPAS was made legitimate to sustain the criminal
responsibility of the accused – in other words: the problems do not lie
in the extent to which the results obtained through COMPAS
contribute, together with the evidential results, to defining the years or
amount of the sentence282.
The application, in all evidence, though developed with the
intention of “solely” calculating the risk of recidivism, has the capacity
to condition the decision-making result: the misgivings raised by
COMPAS do not however finish here.
An analysis carried out by the NGO Propublica on 10,000
defendants has shown how the application is discriminatory: if, on one
hand, it was shown how people of colour are twice as likely to be
classified with a high risk of recidivism, the likelihood is obviously
halved for white defendants283 – raising questions about the software’s
predictive ability.

280
S. QUATTROCOLO, Quesiti nuovi e soluzioni antiche?, cit., 1752; D. KEHL,
P. GUO, S. KESSLER, Algorithms in the Criminal Justice System: Assessing the Use of
Risk Assessments in Sentencing, cit., 18.
281
ECHR, 8th February 1996, John Murray v United Kingdom, ric. n.
19187/91. In this sense, J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 61;
G. CONTISSA, G. LASAGNI, G. SARTOR, Quando a decidere in materia penale sono
(anche) algoritmi e IA: alla ricerca di un rimedio effettivo, cit., 623.
282
Certiorari denied, 137 S. Ct. 2290 (2017)
283
J. ANGWIN, J. LARSON, S. MATTU, L. KIRCHNER, Machine Bias: there’s
software used across the country to predict future criminals. And it’s biased against
blacks, in www.propublica.org, 23rd May 2016; IID., How We Analyzed the COMPAS
Recidivism Algorithm in www.propublica.org, 23rd May 2016, the results o the
analyses are shown below “Our analysis found that:
- Black defendants were often predicted to be at a higher risk of
recidivism than they actually were. Our analysis found that black defendants who

160
 4. Areas of criticism and future prospects

The considerations made so far allow some conclusions in

cost/benefit terms to be drawn for the support that an intelligent
algorithm can bring to the criminal justice system.
Starting from the limits set in the strict domain of the criminal
trial, we will move on to the misgivings that concern a possible
“change of paradigm” of the criminal justice system until arriving at
the “exclusively” technical limits of the algorithms.
A very characterised trial structure emerges from the
combined provisions of Articles 6 and 13 of the European Convention
on Human Rights and 47 of the Charter of Fundamental Rights of the
European Union: it concerns the model of the fair trial, marked by the
principle of cross-examination, and from which a whole set of rights
for the person under investigation/the accused descend as corollaries
– the observance of which is confirmed by the European Ethical
Charter, whose first principle is dedicated to the respect of the

did not re-offend over a two-year period were nearly twice as likely to be
misclassified as higher risk compared to their white counterparts (45 percent vs.
23 percent).
- White defendants were often predicted to be less risky than they
were. Our analysis found that white defendants who re-offended within the next
two years were mistakenly labelled low risk almost twice as often as black re-
offenders (48 percent vs. 28 percent).
- The analysis also showed that even when controlling for prior
crimes, future recidivism, age, and gender, black defendants were 45 percent
more likely to be assigned higher risk scores than white defendants.
- Black defendants were also twice as likely as white defendants to
be misclassified as being a higher risk of violent recidivism. And white violent
recidivists were 63 percent more likely to have been misclassified as a low risk
of violent recidivism, compared with black violent recidivists.
- The violent recidivism analysis also showed that even when
controlling for prior crimes, future recidivism, and age, and gender, black
defendants were 77 percent more likely to be assigned higher risk scores than
white defendants”.

161
fundamental rights284. Among the aspects of a fair trial, those that
impose proper reflection for the entry of “intelligent” algorithms are
probably the principles of the impartiality and independence of the
judge, right of defence and equality of arms.
Under the first profile, it seems fitting to make a distinction
between the modes through which the algorithm’s results enter into
the criminal trial: the reference is to the different phenomena that
pertain to the predictive policing and predictive justice systems.
In fact, in the first case, the algorithms’ results do not have the
power to access the trial directly: they aid the forces of law and order,
but, in the trial, only and always will the facts supporting the charge
and acquired in compliance with the provisions regarding the
acquisition of evidence enter.
The argument for predictive justice is different: the Loomis
case has in fact shown how the calculation of the risk of repeating a
crime has the capacity to condition the judge, thus compromising
his/her function as regards his/her impartiality. In other words: the
impression is that the excessive trust in machines, already widespread,
prevents the judge from being a “virgin mind” for the facts under
his/her scrutiny, prejudicing his/her judgement.

284
European Ethical Charter on the Use of Artificial Intelligence in Judicial
Systems and their Environment, adopted by the (CEPEJ) during its 31st plenary
meeting (Strasbourg, 3-4th December 2018), 7, in www.coe.org: “Principle of respect
for the fundamental rights: ensuring that the processing and implementation of the
artificial intelligence tools and services are compatible with the fundamental rights”;
On this point, S. QUATTROCOLO, Intelligenza Artificiale e giustizia: nella cornice della
carta etica europea, gli spunti per un’urgente discussione tra scienze penali e
informatiche, cit., 4, which states how: “The formula may sound declamatory, but it
must be forcefully underlined that this document, though not binding, is not only and
mainly directed at the states (as often happens for acts of the Council of Europe), but
more at single individuals and, in particular, operators in the private sector, as the
software developers, engineers, mathematicians and even analysts usually are, who
are understandably extraneous to the articulate system of fundamental guarantees
constructed with the Council of Europe”.

162
 The right of defence appears the one most sacrificed by the
“irresponsible” use and development of intelligent algorithms. Under
the first profile, it seems that the use of systems, the operation of which
is not known, precludes, at source, the possibility of exercising the
right of defence effectively: many companies – it must be repeated –
do not provide the source code for the software, because it is covered
by trade secret. This means that the proper “reasoning” of the
algorithm remains precluded to its users285 and, secondly, it could be
theoretically imagined that the programmer favours itself or a circle of
people, preventing the application from processing some data, or
transfers its prejudices into the application286. The possibilities of
neutralising or, even more, of “defending oneself by proving the
contrary” become null, also precluding the person under
investigation/the accused from the possibility of contestation 287. If we
want to make a parallel, the algorithm puts itself as the device’s
reasoning for the output data: a sentence without reasoning, a study
case, is null (both for the Italian system, Article 111, sixth paragraph,
of the Constitution, and in a European perspective, by divesting from
within, the effectiveness of complaint as of Articles 13 of the European

285
V. MANES, L’oracolo algoritmico e la giustizia penale: al bivio tra
tecnologia e tecnocrazia, cit., 14-15, G. CONTISSA, G. LASAGNI, G. SARTOR, Quando
a decidere in materia penale sono (anche) algoritmi e IA: alla ricerca di un rimedio
effettivo, cit., 630; R. BORSARI, Intelligenza Artificiale e responsabilità penale: prime
considerazioni, cit., 264-265; C. CASONATO, Intelligenza artificiale e giustizia:
potenzialità e rischi, cit., 3379-3380; A. SIMONCINI, L’algoritmo incostituzionale:
intelligenza artificiale e il futuro delle libertà, cit., 78-79; F. BASILE, Intelligenza
artificiale e diritto penale: quattro possibili percorsi di indagine, cit., 22-23; O. DI
GIOVINE, Il judge-bot e le sequenze giuridiche in materia penale (intelligenza
artificiale e stabilizzazione giurisprudenziale), cit., 960-961; J. NIEVA-FENOLL,
Intelligenza artificiale e processo, cit., 129-133.
286
Criticisms that, actually, seem anything but theoretical: see supra § 3.1,
note 252 on the predictive policing front; § 3.2, in particular, note 283, for predictive
justice.
287
J. NIEVA-FENOLL, Intelligenza artificiale e processo, cit., 134.

163
Charter of Human Rights and 47 of Charter of Fundamental Rights of
the European Union)288 – it is perhaps precisely because of this that it
would be appropriate to use due precautions in the use of output data,
the processing of which is unknown: the fallibility belongs to men as
it does to machines, the only way to limit the spaces for error is to
permit control over them. This aspect is strictly connected with the
fourth principle of the European Ethical Charter – put in to safeguard
transparency, impartiality and fairness – which invites member states
to “make the data processing methodologies accessible and
comprehensible, authorise external checks”289: in this context, the
pressure to find an appropriate balance between intellectual property
and the requirements of transparency, impartiality and fairness is
strong. On this road, the creation of an ad hoc body, made up of
experts, who properly examine the operation of the software so as to
certify and validate it, could possibly be a first step for attracting these
applications into the orbit of justice – without seeing the right to
defence being sacrificed with it.
On closer inspection, the right of defence seems to be “put into
crisis” under another aspect: systems that operate in a similar way to
COMPAS, where it is the very person under investigation/the accused
who answers some questions, seem to significantly conflict with the
principle of nemo tenetur se detegere, an essential articulation of the
right of defence290, provided for in the domestic and conventional
perspective, by Article 24 of the Constitution and Articles 6 of the
ECHR and 47 and 48 of the Charter of Fundamental Rights of the

288
V. MANES, L’oracolo algoritmico e la giustizia penale: al bivio tra
tecnologia e tecnocrazia, cit., 14-15.
289
European Ethical Charter on the Use of Artificial Intelligence in Judicial
Systems and their Environment, adopted by the (CEPEJ) during its 31st plenary
meeting (Strasbourg, 3-4th December 2018), 11, in www.coe.org.
290
V. MANES, L’oracolo algoritmico e la giustizia penale: al bivio tra
tecnologia e tecnocrazia, cit., 14-15

164
European Union. Under this aspect, it is enough to reflect on the tenor
of the questions placed by the COMPAS questionnaire, which, by
establishing a decimal number capable of influencing the judge’s
decision both on the year and the amount, are placed as a vehicle for
the possible (self)incrimination of the person under investigation/the
accused.
Moreover, on the Italian front, a useful remedy to the adoption
of mechanisms that work in a similar way to COMPAS is represented
by Article 220 of the Code of Criminal Procedure, which expressly
prohibits recourse to expert evidence “to establish the habitualness or
professionalism in the crime, the tendency to commit crimes, the
character and personality of the accused and in general the
psychological qualities independent of pathological causes”.
More generally, the impression is that one is slowly seeing a
change of paradigm: the centre of gravity in criminal justice is
gradually shifting from a criminal law of fact to a criminal law of
author291, to the detriment of the principles of materiality and harm. In
the logic which inspires mechanisms like those examined, it is not that
the fact is reduced to a scenario with unstable outlines, where it is the
author, characterised by his/her personality, to take centre stage,
heralding the return of a positive school – though in version 2.0 –
where you are punished not for what you have done, but for what you
are.
On a strictly technical level, it must be stressed how the
general tendency to place trust in mechanisms, imagined as infallible,
seems to be disproved in several contexts: in fact, machines can not

291
V. MANES, L’oracolo algoritmico e la giustizia penale: al bivio tra
tecnologia e tecnocrazia, cit., 14-15; M. GIALUZ, Quando la giustizia penale incontra
l’intelligenza artificiale: luci e ombre dei Risk Assessment Tools tra Stati Uniti ed
Europa, in www.penalecontemporaneo.it, 28th May 2019, 21.

165
only make mistakes or be biased292, but in the absence of the source
code, nothing can exclude that these mechanisms are designed
according to so-called fuzzy logic293 – i.e. algorithms that are not set
on the binary logic of true/false, but on a more blurred reasoning that
can be summarised as “partially true/partially false”. Under the first
aspect, that the algorithm can implement the prejudices of its
programmer, this is something encountered many times over the
course of this work: the analyses carried out with reference to
COMPAS and Predpol have highlighted its frequency. In these cases,
one sees an indirect prejudice294 which affects the output data from the
algorithm, at the hand of its programmer: the example is the syllogism,
the premise of which, wrong or invalid, is necessarily reflected on the
synthesis. Greater misgivings arise regarding direct prejudice: the
machine proceeds through statistical calculations, processing
countless data, which, with reference to the output can be presented
with greater or lesser frequency. The problem could arise when,
confusing the values between causation and mere correlation 295, the
software makes a connection that is purely random despite the
statistics.
For example, think of software used in calculating recidivism:
the system, in comparing the input data with the dataset, identifies a
link between the place of residence of the person under
investigation/the accused and many convicts who are, subsequently,

292
G. CONTISSA, G. LASAGNI, G. SARTOR, Quando a decidere in materia
penale sono (anche) algoritmi e IA: alla ricerca di un rimedio effettivo, cit., 627.
293
A. SIMONCINI, L’algoritmo incostituzionale: intelligenza artificiale e il
futuro delle libertà, cit., 79.
294
On this aspect, see, in particular, O. DI GIOVINE, Il judge-bot e le sequenze
giuridiche in materia penale (intelligenza artificiale e stabilizzazione
giurisprudenziale), cit., 958-959; C. CASONATO, Intelligenza artificiale e giustizia:
potenzialità e rischi, cit., 3376-3377.
295
G. CONTISSA, G. LASAGNI, G. SARTOR, Quando a decidere in materia
penale sono (anche) algoritmi e IA: alla ricerca di un rimedio effettivo, cit, 627.

166
recidivists; in these cases, the system could show direct prejudice,
confusing mere correlation with causation.
The only useful tool to confront these criticisms is to define a
legal remedy that at least allows the potentially fallible results to be
neutralised.

167
 CHAPTER V

DATABASES AND PREDICTABILITY.

THE FINANCIAL RELATIONS ARCHIVE AND THE RISK ANALYSIS FOR
EVASION

Pietro Sorbello

Table of Contents: 1. Foreword. 2. Police activity, the right to the protection of

personal data and the general interest of preventing and pursuing crimes
(hints). 3. The thread between economic analysis of law and criminal policy,
new technologies and criminal law (and procedure). 4. The evolution of the
regulation of the financial relations archive. Analysis of the risk of evasion
and the interoperability of databases. 5. The acquisition of notices of
“theoretical” crime. Final reflections

1. Foreword

In a hypothetical dystopian society regulated by illiberal

criminal law, as outlined in 1956 by Philip Dick in his story “The
Minority Report”, crime prevention even arrives at inflicting the
criminal penalty for a crime that has not yet been committed, but has
been predicted. In fact, according to this predictive model, even the
mere intention constitutes a crime and directly opens the prison doors
or, in the best case scenario, permits accused persons, with a Kafka
type unawareness, to be subjected to trial.
If the principle of materiality provides an insuperable limit to
criminal responsibility and the re-education aims do not allow the
individual to be instrumentalised for the general purposes of criminal
policy, predictivity cannot easily be reconciled with the idea of

168
criminal justice296, in the strict meaning of a judgement on criminal
responsibility based on the judge’s free conviction. However,
significant margins of use for predictive algorithms remain in
enhancing criminal proceedings, but also (and perhaps especially) in
the prevention of crimes.
The use of predictive algorithms in police security activities,
aimed at preventing deeds contrary to the legal system or in any case
ones capable of breaking orderly safe social cohabitation, in fact
makes previously unimaginable computation capabilities available for
analysing and comparing data, obtaining a synthesis of complex
information to direct and optimise police activity (decision making);
one can thus place the analysis of metadata, i.e. the relationships
between data singularly or ones that are not immediately significant,
alongside the traditional information activities based on observation
and deductions.
The new technologies provide an undeniable contribution in the
prevention and acquisition of notice of crime, which represent the first
contact between the law and criminal procedures, marking the passage
from the powers of administrative police to the functions of judicial
police. This contribution is particularly significant in the analysis of
the risk of evasion, and the observance of the parameters of purpose

296
In sentence No. 313 of July 1990, the Constitutional Court stated that if the
finalisation of the sentence was directed towards the functions of social defence and
general prevention, instead of the principle of re-education, “one would run the risk
of instrumentalising the individual for the general purposes of criminal policy (general
prevention) or favouring the satisfaction of the collective needs for stability and
security (social defence), sacrificing the individual through the exemplary nature of
the penalty. It is for this reason that, in an evolved state, the aim of re-education cannot
be considered, extraneous to the very legitimation and function of the penalty”. In
Foro it., 1990, I, 2385, with note by G. FIANDACA, Pena “patteggiata” e principio
rieducativo: un arduo compromesso tra logica di parte e controllo giudiziale.

169
and proportionality297 make the use of them lawful298 when finding a
balance between general interests and the protection of personal data.

2. Police activity, the right to the protection of personal

data and the general interest of preventing and pursuing crimes (hints)

The concept of security has its roots at the origins of all forms
of social cohabitation299.
Ever since the individual associated with others, the need to
safeguard the survival of not just individuals but also their group
against possible internal and external dangers was born: the legal
phenomenon thus takes on importance as an inevitable set of rules
needed for the peaceful cohabitation of the collective.

297
The reference is to Article 52 of the ECHR, by which any limitations to the
exercise of the rights and freedoms recognised by the Charter, in particular the right
to the protection of personal data (Article 8), “must be provided for by the law and
respect the essential content of said rights and freedoms. In compliance with the
principle of proportionality, limitations can only be brought where they are necessary
and effectively meet the purposes of general interest recognised by the Union or the
requirement to protect the rights and freedoms of others”.
298
Under Article 5, first paragraph, of Legislative Decree No. 51 dated 18th
May 2018, implementing EU Directive 2016/680 of 27th April 2016, on the protection
of natural persons with regard to the processing of personal data by competent
authorities for the purposes of prevention, investigation, fact-finding and the
prosecution of crimes or the execution of criminal penalties, “the processing is lawful
if it is necessary for the execution of a task of a competent authority for [such]
purposes and it is based on European Union law or provisions of law or, in the cases
provided for by law, the regulations that identify the personal data and the purposes
of the processing”.
299
Thus, T. GIUPPONI, Sicurezza personale, sicurezza collettiva e misure di
prevenzione. La tutela dei diritti fondamentali e l’attività di intelligence. Intervention
at the “Collective Security and Fundamental Rights” seminar held at the University of
Ferrara on 26th September 2007, www.forumcostituzionale.it.

170
 If security became the monopoly of the absolute sovereign 300
with the birth of the state, it is with its evolution to the liberal state of
law that the pre-requisites for a modern vision of the relationship
between law, freedom and government 301 are declared in a clean break
from the ideological and political substrate of the absolute regime,
where the individual is at the centre of the legal system and his/her
rights are solemnly declared302. The relationship between security and
rights has its origin in this context, it follows the parallel evolution of
the concept of state and legality and incorporates constitutionalism,
where the safeguard of freedom is not exhausted in the non-
interference of the state, because the latter must guarantee individual
rights and foster the person, both as an individual and in social
formations: with the advent of so-called positive freedom, the concept
of security, originally anchored to the guarantee of ne cives ad arma
veniant, and the minimal conditions of life in common changes 303.

300
The purpose of the state is thus “the peace and defence of all and anyone
who has right to end has the right to the means”, including the sovereign power “to do
everything that it thinks necessary to be done, both in advance to preserve the peace
and security, predicting discord within and hostility without, and to re-acquire peace
and security when they are lost” Thus, T. HOBBES, Leviatano (trad. G. MICHELI),
Rizzoli, 2011, 177.
301
Thus the general principles of legal reservation for the limitations on liberty
and ownership, political representation and separation of powers take form: “wherever
two men who cannot appeal to a certain law and a common judge on earth to decide
the legal disputes between them, in this case one is still in a state of nature [...] and
therefore also every absolute principle for those who are under its dominion”. Thus,
J. LOCKE, Due trattati sul governo (trad. B. CASALINI), Pisa University Press, 2007,
239 et seq. For the separation of powers, see C.L. MONTESQUIEU, Lo spirito delle leggi
(translated by B. BOFFITTO SERRA), Rizzoli, 1967, 207.
302
See, for example, the declaration of Independence of the American
colonies of 4th July 1776 and the Declaration of the Rights of Man and of the Citizen
adopted on 26th August 1789.
303
Protection of liberty and guarantee of security connote the state in an
instrumental function causa hominum: “laws are the conditions with which individual
isolated men join together in society, tired of living in a continuous state of war and
enjoying a liberty made useless by the uncertainty of preserving it. They sacrifice a

171
 To this end, the system provides for “preventive and repressive
measures aimed at maintaining public order, meaning all those
fundamental legal assets and primary public interests on which orderly
civil cohabitation is supported in the national community, as well as
the security of the institutions, citizens and their assets”304.
The further relevant concept is security. Apart from its
minimum interpretation coinciding with physical safety, it describes
the situation where the citizens are assured “the peaceful exercise of
those rights of liberty that the constitution guarantees with such
strength. One has security when the citizen can go about his/her lawful
business without being threatened by harm to his/her physical and
moral personality; it is orderly civil living, which is undoubtedly a half
of a state of law, free and democratic305.
The purpose of the public security police is strictly connected
with the protection of public order and security. This is identified by
Article 1 of Royal Decree No. 773 of 18.06.1931306, under which “the

part of it to enjoy the remaining part in security and tranquillity and it was therefore
necessity that forced men to cede part of their liberty, but no-one wanted to put it in
public deposit unless it was the minimum portion possible, just enough to induce
others to defend it”. C. BECCARIA, Dei delitti e delle pene, in L. FIRPO (diretto da),
Edizione nazionale delle opere di Cesare Beccaria, Mediobanca, 1984, 25.
304
Thus, Article 159, second paragraph, of Legislative Decree No. 112 of 31st
March 1998 on the transfer of administrative functions and tasks from the state to local
bodies. The concept of public order is therefore given by “those fundamental legal
assets and primary public interests of orderly civil cohabitation is supported in the
national community. The protection of those interests, which include the physical and
psychological integrity of people, the security of possessions and the observance or
guarantee of every other legal asset of fundamental importance for the existence and
performance of the system, represents the core of the public security police functions
[attributed exclusively to the state”. In these terms, Constitutional Court ruling No.
218 of 11th February 1988.
305
In these terms, Constitutional Court No. 2 of 23rd February 1956 in Riv. it.
dir. pen., 1956, 441, with note by P. NUVOLONE, Appunti e spunti tra precetti e
sanzioni.
306
On the approval of the Consolidated Act on the public security laws.

172
public security authority watches over the maintenance of public
order, the security of the citizens and their safety and the safeguard of
property”. In the perspective of prevention, the task of the state “is not
so much (or better, is not only) to guarantee the right to the personal
security of the single individual as the overall security of the citizens’
rights and the legal assets implied in them, in a complex social
context”307. This dual dimension of security in effectively synthesised
in Article 2 of the Constitution and subsequent Articles 13, 14 and 15,
which, in recognising and guaranteeing the inviolable rights of man,
nevertheless allows limitations through deed reasoned by the judicial
authorities and in the cases strictly provided for by the law308.
Alongside the so-called “negative liberties”, an expression of
the classical liberal concept dealing with the protection of the person
against undue limitations by public authorities, allowed by the dual
reservations of law and jurisdiction (e.g. Articles 13, 14 and 15 of the
Constitution), with the affirmation of the liberal/democratic
constitutionalism and the imposition of solidarity, the state also

307
Thus, T. GIUPPONI, Sicurezza personale, cit., for whom this vision “puts
not so much the individual, meaning a being in itself isolated from the reference social
context, but the human being in its social relations, the centre of the apportionment of
rights and duties [to guarantee] not just “security from” potential intrusions in the
range of individual liberties, but also the “security of” being able to express its
personality in full through the constitutional heritage of rights and in the context of
(and not disregarding) the reference social context”. In an accordant sense, for the
State Constitution, Sect. VI, ruling No. 85 of 16.01.2006, “individual liberties, in the
social state, must be married in unison with the interest of the collective and undergo
compressions in some modes of exercise to make them compatible with the equal
orderly liberties others”.
308
This ambivalence is also found at international level, e.g. Article 8, second
paragraph, of the ECHR on the right to respect for private and family life, under which
“There shall be no interference by a public authority with the exercise of this right
except such as is in accordance with the law and is necessary in a democratic society
in the interests of national security, public safety [...] the prevention of disorder or
crime [...]” See also the cited Article 52, first paragraph, of the Charter of Fundamental
Rights of the European Union.

173
intervenes in the so-called “positive liberties”, concerning social and
economic rights (e.g. Articles 32, 41 and 42 of the Constitution).
Thus, the individual and collective requirements for protection
come together and, in this perspective, the extended interpretation of
the notion of security allows both the specific economic/financial
profile, in which economic liberties can be exercised, and the
institutional mission of the Guardia di Finanza [Financial Police] as a
police force, which exercises solely the functions of
economic/financial policing, to be upheld309.
More generally, the security police materialises in the various
activities aimed at finding and searching for “objective” situations of
danger or the start of criminal activities. It differs from the
administrative police which, following the transfer of administrative
functions from the state to the regions and local authorities, has
become instrumental for the active administration functions 310, and
also from the judicial police, the functions of which can be exercised,

309
Under Article 2 of Legislative Decree No. 68 of 19th March 2001 “1. [...]
the Financial Police Corps carries out economic and financial policing functions to
protect the public budget and that of the regions, local authorities and the European
Union”. For an in-depth examination, see C. RICOZZI – N. DI PAOLO, La funzione di
polizia economica e finanziaria, in Riv. Guardia di Finanza, 5/2003, 1485.
310
“The apportioning of the powers for police functions is founded on the
distinction between the jurisdictions regarding public security, reserved exclusively to
the state under Article 4 of Decree of the President of the Republic 616/1977, and the
others set out by the broad category of administrative police and transferred to the
regions as functions accessory to the material sectors attributed to them. Whilst the
first [...] concern the preventive and repressive measures aimed at maintaining public
order and, therefore, refer to the activities traditionally included in the concepts of
judicial police and public security police (in the strict sense), the others concern the
activities of prevention and repression aimed at preventing damage or injury that could
be brought to persons or things in the execution of activities included in the subjects
over which they exercise regional jurisdiction [...], without the assets or interests
protected in the name of public order being harmed or put in danger”. Thus,
Constitutional Court No. 218 of 11th February 1988.

174
under the control and management of the judicial authorities, as
evidence of a criminally significant deed arises 311.
If the balancing logic justifies the shrinking of the individually
protected interest, according to the same assessment method for the
concrete implementation of justifications, in the broadest meaning of
security, a routine implicit limit is connected with the prevention
activities supplementing police activity outside the cases which make
the legal jurisdictional reservations necessary: in fact, the police
powers constitute an external limit in differentiated forms because
there are different police powers.
For our purposes, we will consider the use of databases and the
balancing between the right to the protection of personal data provided
for by Article 8 of the Charter of Fundamental Rights of the European
Union and the general interest for the prevention, investigation,
detection and prosecution of crimes, as governed by Legislative
Decree No. 51 of 18.05.2018 implementing Directive 2016/680 of
27.04.2016312. In its preamble, the Directive provides that:

- “The activities of the police [...] mainly revolve around the

prevention, investigation, detection and prosecution of crimes,
including police activities carried out without prior knowledge of the

311
The link is in Article 220 of the provision implementing the Code of
Criminal Procedure, under which “when evidence of crime emerges during the
inspection and supervisory activities provided for by laws or decrees, the deeds
necessary to assure the sources of evidence are carried out in compliance with the
provisions of the code”. These deeds fulfil the judicial police functions governed by
Article 55 of the Code of Criminal Procedure.
312
On the discipline for this specific area of processing. In fact, under Article
1, second paragraph, of Regulation 2016/679/EU of 27th April 2016 on the protection
of natural persons with regard to the processing of personal data, “This Regulation
does not apply to the processing of personal data: [...] d) by competent authorities for
the purposes of the prevention, investigation, detection or prosecution of criminal
offences or the execution of criminal penalties, including the safeguarding against and
the prevention of threats to public security”.

175
criminal significance of a deed. These activities [...] also include the
maintenance of public order as a task conferred on the police [...]
where necessary to safeguard against and to prevent threats to public
security and the fundamental interests of a society protected by the law
which can give rise to crimes” [...] (No. 12);
- “it is necessary for the competent authorities to process the
personal data collected for the purposes of preventing, investigating,
detecting or prosecuting specific crimes outside such context to
develop knowledge of criminal activities and to connect the different
crimes detected” (No. 27).
This last point includes and explains the purposes and need for
the automated processing of personal data by the police. This is made
possible by new technologies that allow data to be analysed and
compared, returning a synthesis of complex information to direct the
prevention activities: the traditional search by investigators for notices
of crime is now also accompanied by the analysis of the metadata, i.e.
the system of relations between individually insignificant data. Since
Article 1, second paragraph, applies “to the fully or partially
automated processing of the personal data of natural persons [...]
carried out by the competent authorities for the purposes of the
prevention, investigation, detection and prosecution of crimes or the
execution of criminal penalties, including the safeguard against and
protection from threats to public security”, we need to stress the impact
of the new technologies on the information system of the judicial
police, whose activity is now eased by the consulting of databases and,
especially, their interoperability.
This analysis will therefore by set out with reference to tax
evasion, integrating the infringement of the tax obligation, as an
expression of the duty of economic solidarity313.

313
Given that a fact expressing the contributory capacity is a fact that
expresses economic strength, “connecting taxes to the contributory capacity [Article

176
 3. The thread between economic analysis of law and
criminal policy, new technologies and criminal law (and procedure)

Among the possible interrelationships between the law and

economics, the economic analysis of law studies the effectiveness of
the legal regulations and assesses what behaviour they
(dis)incentivise. Since it considers these regulations as a government
tool to reach determined objectives, the economic analysis method
applied to the criminal policy choices allows a rational rethinking of
the penalty system314.
From the perspective of a generalisation, within the limits of the
possible, of this method, the scientific usefulness of the theory of the

53 of the Constitution], this must be seen, from the legislator’s perspective, as a

regulation that binds it and limits its power and, from the contributors’ perspective, as
a regulation of guarantee”. Thus, F. TESAURO, Istituzioni di diritto tributario,
Giappichelli, 2011, 65. The connection between Articles 2 and 53 of the Fundamental
Charter has been identified in constitutional case law. See Constitutional Court, 3rd
February 1992, No. 51, in Riv. dir. trib., 2/1992, 566, with note by G. FALSITTA,
Epicedio per il segreto bancario nei confronti del Fisco. Nevertheless, in this
relationship, the vertical profile of economic solidarity, that of the individual to the
public administration, rather than the horizontal profile between and to the members
of the collective, has become progressively emphasised. On this subject,
Constitutional Court, 12th July 2000, No. 351.
314
Actually, the use of this tool has been limited to economic crimes
characterised by the “computer hacking model” and would not be applied to crime for
non-financial reasons. Thus, F. MANTOVANI, La “perenne crisi” e “la perenne
vitalità” della pena. E la “crisi di solitudine” del diritto penale, in E. DOLCINI -
C.E.PALIERO (edited by), Studi in onore di Giorgio Marinucci, I, Giuffrè, 2006, 1173.
Nevertheless, the applicability of the method appears to exceed the (objective) context
of the “criminal law of economics” to turn to the (subjective) profile of the rational
agent who remains such even if it does not pursue an advantage, but only wants to
remove or mitigate a disadvantage, directing its choices in relation to all the negative
consequences. Please refer to P. SORBELLO, Politica criminale ed osservanza delle
regole. Riflessioni su limiti e possibilità di conversione al razionale dei
comportamenti, in Riv. it. dir. proc. pen., 4/2017, 1934.

177
economically based crime315 must be verified taking into account that,
alongside the “modern” institutes of criminal law (e.g. the
responsibility of organisations and confiscation by equivalent), the
economic analysis of the law also provides useful indications in
tackling the traditional categories, with the objective of intervening on
the more significant regulatory problems. In order to guarantee the
rationality (and hence the effectiveness) of the crime system, the crime
policy must intervene on the probabilities of being discovered, of the
penalty being applied and of the cancelling of any profit obtained from
the breach of the rule: since economic analyses valorise the connection
between crime policy strategies and assessment, for the potential
aggressor, of the costs and benefits linked to breaching the rule, the
rational person’s propensity for breaching is inversely proportional to
these probabilities.
The reflections below will only consider the first of the
regulatory problems.
Since increasing the probabilities of discovering criminal acts
encounters structural limits, which do not allow the public control
authorities to be strengthened without limit, a significant contribution
in acquiring a notice of crime and, at the same time, on the number of
unrecorded crimes comes from the new technologies. Although the
judicial police’s information system, the first of those governed by
Article 55 of the Code of Criminal Procedure, hardly appears
interesting for criminal law, it is however particularly significant for
criminal law in the perspective (of crime policy) of limiting crimes in
the criminal system, also by means of tools that are not of a criminal
nature or exclusively sanctionative.
The thread that links the economic analysis of criminal law and
policy, new technologies and criminal law (and procedure) is now less

315
Thus, C.E .PALIERO, L’economia della pena (un work in progress), in Riv.
it. dir. proc. pen., 4/2005, 1396.

178
clear because a solution is offered by the new technologies,
particularly by the automated processing316 of (personal) data “for the
purposes of the prevention, investigation and detection of crimes”, as
provided for Legislative Decree 51/2018.
However, the opportunities offered by the new technologies are
accompanied by the risk of decisions based solely on automated
processing (including profiling) which produce negative effects for the
data subject: it is an insurmountable red line, guaranteed by the
prohibition set out by Article 11 of said Directive317 and bringing with
it the ethical implications of automation regarding the discriminatory
consequences.
These aspects, which go beyond the horizons of our reflections,
are in any case a warning in remembering that “the processing of
personal data should be designed to serve mankind” 318 and inspire a

316
Under Article 2 (Definitions) of Legislative Decree 51/2018, this means
“[...] b) processing: any operation of set of operations, carried out with or without the
aid of automated processes, applied to personal data or sets of personal data, such as
collection, recording, organising, structuring, preservation, adaptation or
modification, extraction, consultation, use, communication by transmission,
dissemination or any other form of making available, comparison or interconnection,
limitation, deletion or destruction; [as well as for] e) profiling: any form of automated
personal data processing consisting of the use of such data to assess determined
personal aspects for a natural person, in particular to analyse or predict aspects
regarding [...] the economic situation [...]”.
317
For Consideration No. 38, “the data subject must have the right not to be
the subject of a decision which evaluates personal aspects concerning him/her solely
on the basis of automated processing and which produces an adverse legal effect
concerning the data subject or significantly affects him or her. [...] Profiling that
results in discrimination against natural persons on the basis of personal data which,
by their nature are particularly sensitive under the profile of fundamental rights and
freedoms must be prohibited [...]”. This prohibition is provided for by Article 8 of
Legislative Decree 51/2018.
318
Thus Consideration No. 4 of EU Regulation 2016/679, which however
specifies that “the right to the protection of personal data is not an absolute right; it
must be considered in relation to its function in society and be balanced against other
fundamental rights, in accordance with the principle of proportionality”.

179
use of new technologies “in the direction of a new digital
humanism”319.

4. The evolution of the regulation of the financial

relations archive. Analysis of the risk of evasion and the
interoperability of databases

If the punishment is the tool prepared by the system to ensure

that the precept is observed, this has further importance at a taxation
level, because the function of ensuring compliance with the precept is
accompanied by the one of guaranteeing the fiscal levy.
In a tax system based on the spontaneous fulfilment of the tax
obligation, the ability of the regulation to direct behaviour according
to expectations is subordinate to the message that unlawful tax savings
is not only never convenient, but is also expensive: “an effective anti-
evasion policy is based on the deterrent capacity of the administration

319
In these terms, A. SORO, Proteggere i dati per governare la complessità.
Speech of the Chairman of the Guarantee Authority for the protection of personal data,
Rome, 2017, 8 et seq., for whom “the intelligence of their choices depends on the
exactness of the data used and the processing logic on which the algorithm
configuration is based. [...] The possible implications, on a social scale, are anything
but marginal. Algorithms are not neutral calculation syllogisms, but human opinions
structured in a mathematical form which, as such, reflect, to a more or less significant
extent, the pre-comprehensions of those who design them, risking turning algorithmic
discrimination into social discrimination. With regard to these risks, the guarantees
sanctioned by the new legal framework for automated decision-making processes are
important, ensuring [...], at least in the final instance, the filter of man to combat the
unconditional delegation to the blind determinism of technology. [...] Beyond the
almost primaeval fear of mankind being the victim of its creation, there thus emerges
the need to lay down solid ethical and legal bases for a truly sustainable development,
because technology must be able to serve and supplement human intelligence without
replacing it”.

180
rather than an unimplementable and costly check on the mass of
contributors320.
Since the choice to evade depends critically on the perceived
probability of being subject to a tax investigation, a watershed in
increasing the control capacity of the financial administration took
place with Articles 18 and 20 of Act. No. 413 of 30.12.1991321. In
particular, under the fourth paragraph of Article 20, “with decree of
the Treasury Ministry in agreement with the Ministries of the Interior
and Finance, the destination and methods of communications, with the
maximum confidentiality, by companies and credit institutions and the
post office administration as well as trust companies and every other
financial broker for the identification data, including the tax code, of
every entity that has accounts or deposits with them or may in any case
dispose of them, as well as the criteria for their use, are established.
This is to be issued within sixty days of the act coming into force”.
By interrogating the tax register of account and deposit
relationships, it would therefore be possible 322 to know, with reference

320
“From a theoretical perspective, people evade because an economic
convenience exists. [...] if the benefit arising from evasion exceeds the expected cost
associated with the possibility of detection”. Thus, the Ministry of Economics and
Finance, I risultati della lotta all’evasione. Relazione al Parlamento, Roma, 2007,
par. 16.
321
On, inter alia, provisions to facilitate and enhance the detection activity.
322
Despite the term of sixty days, the register of account and deposit
relationships was set up by Ministerial Decree No. 269 of 04.08.2000. On this subject,
B.D. TRIDICO, L’utilizzo dell’anagrafe dei rapporti finanziari ai fini dell’attività di
controllo fiscale, Relazione Corte dei Conti – Sez. centrale di controllo sulla gestione
delle Amministrazioni dello Stato (26th July 2017, No. 11/2017/G) censured that “the
decree was adopted after ten years and, however, has never really been concretely
implemented. In reality, the register of financial relationships only actually became
operative and accessible by all qualified persons in 2009. If therefore particularly
important delays emerged for its final establishment, much more serious is the
situation encountered for its effective concrete use to fight evasion, for which serious
dereliction must be noted for the Revenue Office, which never prepared the selective
lists provided for or, later, the evasion risk analyses [...]”. The provisions for the

181
to specifically identified natural persons or legal entities, any existence
of account or deposit relationships under their name or joint name or
for which they hold in the name and on behalf of someone else.
In starting the progressive dismantling of banking secrecy, the
new discipline made banking investigation available “for carrying out
tax activities provided for by the law, detection aimed at the search for
and acquisition of evidence and sources of evidence during criminal
proceedings, both in the preliminary investigation stages and in the
subsequent trial stage, i.e. assessments of an asset nature for the
purposes of prevention provided for by specific provisions of law and
for the application the preventive measures”323.
Compared to the original provision of Act 413/1991, the bank
investigations and the register of accounts and deposits have been
affected by significant amendments over time, including:

- Paragraphs 402 and 403 of Article 1 of Act No. 311 dated

30.12.2004 (2005 financial act) which, in the context of the respective
sections on income tax and value added tax, extended the investigation
powers on both a subjective level, with the financial administration
being able to request any information from any credit or financial
operator (including trust companies, however they may be called) and
an objective level, with regard to any type of relationship or
transaction (in the all-embracing sense): this enhancement marked the
passage from bank investigations to financial ones.
- Article 37, fourth paragraph, of Law Decree No. 223 of
04.07.2006324, which implemented the tax register system by

evasion risk analyses were issued with order No. 197357 of 31st August 2018 by the
Director of the Revenue Office.
323
Thus the first paragraph of Article 4 of Ministerial Decree 269/2000.
324
On urgent provisions, inter alia, regarding revenue and tax evasion,
converted with amendments by Act No. 248 of 04.08.2006.

182
providing for the obligation of periodic notices from financial brokers:
thus, the automatic identification of brokers in relation to the
contributor became possible for the financial administration 325.

This article set up the register of relations between financial

brokers and contributors, as a section of the tax registers, in
compliance with Article 7 of Decree of the President of the Republic
No. 605 of 29.09.1973326, fed by the monthly notices for the existence
and type of relations 327, with indication of the personal details of the
holders, including the tax code. Again, Law Decree 223/2006
established the general rules for access to the archive, providing that it

325
The apron of the intended addressees was broadened by Law Decree no.
98 of 06.07.2011, converted with amendments by Act No. 111 of 15.07.2011, with
the insertion of insurance brokers, limited to the financial activities carried out.
326
On the provisions for the tax register and the tax code of the contributors.
327
It also includes the information on repatriated financial availabilities
observing the procedure as of Article 13-bis of Law Decree No. 78 of 01.07.2009,
converted with amendments by Act No. 102 of 03.08.2009 (so-called “tax shield ter”).
With the provision of 9th August 2013, the Revenue Office specified “since all the
accounts will arrive in an indistinct manner, it seems consistent that the shielded
accounts are also included in the flow”

183
could be consulted for strict purposes 328, beyond investigations in the
tax field which have been progressively broadened over time329.

328
Under Article 7, eleventh paragraph, of Decree of the President of the
Republic 603/1975, the notices are used:
- for the requests as of Article 32, first paragraph, number 7), of Decree of the
President of the Republic No. 600 of 29th September 1973 and Article 51, second
paragraph, number 7), of Decree of the President of the Republic No. 633 of 26th
October 1972, on the assessment of direct and value added taxes;
- for the activities connected with collection through role;
- by the persons as of Article 4, paragraph 2, letters a), b), c) and e) of
Ministerial Decree No. 269 of 4th August 2000, for the assessments aimed at
searching for and acquiring evidence and sources of evidence during criminal
proceedings, both for the purposes of the preliminary investigations and exercising
the functions provided for by Article 371-bis of the Code of Criminal Procedure, and
for the assessments of an asset nature for the purposes of prevention provided for by
specific provisions of law and for the application the preventive measures.
329
Access to the archive of financial relations is also provided for by:
- Article 187-octies, fourth paragraph, letter e-bis), of legislative Decree No.
58 of 24th February 1998, with reference to the powers of CONSOB on market abuse;
- Articles 1, fifth paragraph, letter b), of Law Decree no. 262/2006 of 3rd
October 2006 and 15, paragraph 8-duodecies, of Law Decree No. 78 of 1st July 2009,
with regard to the administrative and tax duties of the Customs and Monopolies
Agency;
- Article 8 (now 9) of Legislative Decree no. 231 of 21st November 2007, in
the exercise of the police monetary powers on money-laundering;
- Article 83, paragraph 28-sexies, of Law Decree No. 112 of 25th June 2008,
for consultation by Municipalities and the persons authorised by them to collect taxes;
- Article 25 of Law Decree No. 83 of 22nd June 2012, extending the monetary
powers in checking public expenditure;
- Article 19 of Law Decree No. 132 of 12th September 2014, which added
article 492-bis of the Code of Criminal Procedure: subject to authorisation of the court,
the creditor is guaranteed that the information contained therein can be used for a
better efficiency of the executive process;
- Article 3 of Law Decree No. 193 of 22nd October 2016, which allowed the
Revenue Office and its collection body to also use the archive for the purposes of the
collection functions.

184
 - Article 63 of Law Decree No. 231 of 24.11.2007330, with
provision of the obligation to communicate the information on any
financial transaction implemented by its clients outside an ongoing
relationship, together with the identification data and tax code of the
persons who have any relationship or carry out transactions outside an
ongoing relationship on its own behalf or on behalf of others;
- Article 11, fourth paragraph, of Law Decree No. 201 of
16.12.2011331, which overcame the approach for financial
investigations by which the contributor to be checked had first to be
identified and only afterwards could the archive of financial relations
be consulted. In fact, on the basis of such provision, all the information
communicated by the financial operators is used by the Revenue
Office “for the identification of the contributors at greatest risk of
evasion which should be checked”332.
- Paragraph 314 of Article 1 of Act No. 190 dated 23.12.2014
(2015 stability act), which intervened on the above-mentioned Article
11, fourth paragraph, of the “Save Italy” Decree, perhaps marked the
most interesting amendment for the purpose of these reflections.

330
Implementing Directive 2005/60/EC on the prevention of the use of the
financial system for the purpose of money laundering and terrorist financing.
331
Implementing the urgent provisions for the growth, equity and
consolidation of public accounts (so-called “Salva Italia [Save Italy]” Decree),
converted with amendments into Act No. 214 of 22nd December 2011.
332
From 1st January 2012, financial brokers were obliged to periodically
communicate to the tax register not only the handling that concerned their ongoing
relationships, the mere existence of which was already subject to communication, and
all other necessary information referring to these relationships for the purposes of tax
controls, but also the amount of the extra-account transactions (request for cash
through cheques and bank orders, currency exchange and cheques, etc.), for which
provision had already been made for some time regarding the transmission of the
personal details of the holders and the persons carrying them out. As stressed in the
report accompanying the decree, the reasoning for the intervention was to acquire the
handling of financial data periodically and not just for the checking activities that had
already been started, thus providing the financial administration with an
extraordinarily effective tool in combating tax evasion.

185
 If the Revenue Office could use all the information in the tax
register to draw up selective check lists before the 2015 Stability Act,
the same information, including the average annual value of the
relationships’ deposits333, could afterwards be used for evasion risk
analyses, in accordance with the criteria set annually by the financial
administration334.
- Paragraph 14 of Article 9-bis of Law Decree No. 50 dated
24.04.2017335 provided that “the Revenue Office and the Financial
Police, when defining specific control strategies based on the analysis
of the tax evasion risk, take into account the fiscal reliability level for
contributors arising from the application of the indices and
information in the special section of the tax register as of Article 7,
sixth paragraph, of Decree of the President of the Republic No. 605
dated 29.09.1973”;
- Article 16-quater of Law Decree No. 119 dated 23.10.2018 336,
intervening again on Article 11 of the “Save Italy” Law Decree,

333
The same information is also used for the purposes of simplifying the
citizens’ performances with regard to filling in the single substitute declaration as in
Article 10 of the regulations contained in Decree of the President of the Council of
Ministers No. 159 of 5th December 2013, regarding the methods of determination and
fields of application of the equivalent economic situation indicator (ISEE), as well as
at the time of checking the truthfulness of the data declared in that declaration.
334
The criteria were adopted with the provision of the Director of the Revenue
Office No. 197357 of 31st August 2018, which provided for an experimental
procedure for analysing the risk of evasion. This phase, regarding the tax position of
associations of persons and capital companies, consists of the integrated use of the
information communicated by the operators to the financial relations archive and the
other elements in the tax register and will allow the extrapolation of the companies for
which, though in credit on their current accounts, for tax year 2016, the declaration
for the purposes of direct taxes and VAT was omitted or submitted without any
significant accounting data.
335
On urgent provisions for finance converted with amendments by Act No.
96 of 21st June 2017.
336
On urgent provisions for taxation converted with amendments by Act No.
136 of 17th December 2018.

186
provided that the provision with which the Director of the Revenue
Office establishes the modes for the periodical communication to the
tax register, includes “adequate security measures, of a technical and
organisational nature, for the transmission of the data and their storage,
which cannot exceed ten years”.
The same article also provided for the use, by the Financial
Police, of the information contained in the financial relationships
archive to analyse the risk of evasion, thus allowing more targeted
investigations, based on the exchange of information with the Revenue
Office, including that received in the context of the exchange of
information provided for by Directive 2011/16/EU of 15.02.2011 337
and agreements between Italy and foreign countries.
The horizon of the exchangeable information in the context of
administrative cooperation was significantly broadened with Directive
2016/2258/EU of 06.12.2016338, which allowed the tax authorities
access to the money-laundering information339.

337
Regarding administrative cooperation in the tax sector, which has allowed
the financial administration to have further information resources available, coming
from similar administrations in the member states of the European Union, through a
differentiated information exchange system. In fact, under Legislative Decree No. 29
of 4th March 2014, the exchange can take place on the basis of a specific request
(Article 4), automatically for an information set to be communicated compulsorily
(Article 5) or willingly (Article 6).
338
On the amendment to Directive 2011/16/EU regarding access by the tax
authorities to the information on money-laundering.
339
Implemented with Legislative Decree No. 60 of 18th May 2018, which in
intervening on Article 3 of Legislative Decree 29/2014 provided that:
- “the linking services [...] provide the other member state’s requesting
authority with all the elements useful for the exchange of information and
administrative cooperation. To this end, they use the data and notices, acquired under
Decree of the President of the Republic No. 605 of 29th September 1973, and have
access to the data and information on the actual ownership of legal entities and trusts
contained in a special section of the Companies’ Register, as per Article 21 of
Legislative Decree No. 231 dated 21st November 2007, using the methods as of

187
 This last profile expresses the synergistic and “bidirectional”
importance340 of the cross-matching of information and, through that,
the effectiveness of the system to prevent and combat tax evasion and
laundering of illicit gains. If the archive of financial relationships can
be consulted for the in-depth investigation of reports of suspicious
transactions (Article 9, sixth paragraph, of Legislative Decree
231/2007), the exchange is bidirectional because the money-
laundering discipline makes all the information acquired by the
Financial Police in the context of the inspections and in-depth
investigations of reports of suspicious transactions (Article 9, ninth
paragraph) and that stored by the obliged entities (Article 34, first
paragraph) usable for tax purposes.
The last intervention that occurred was with Article 1,
paragraphs 681 to 686, of Act No. 160 dated 27.12.2019 341, which
pursued the objective of combining the protection of personal data 342

paragraph 2, letter d) and paragraph 4, letter c) of the same article. […]” (third
paragraph);
- for the purposes of “administrative investigations as of paragraph 3, in the
context of exercising the powers provided for by Title IV of Decree of the President
of the Republic No. 600 dated 29th September 1973, the Revenue Office and the
Financial Police are allowed access to the documents, data and information acquired
in fulfilling the obligation to make adequate checks on the customers under Article 18
of Legislative Decree No. 231 of 21st November 2007, using the methods as of Article
19 of the afore-mentioned legislative decree, and stored under Article 31 with the
methods as of Article 32 of the same legislative decree” (comma 3-bis).
340
For in-depth analysis, see M. CARBONE, Archivio dei rapporti finanziari:
esigenze di razionalizzazione tra normativa fiscale e antiriciclaggio, Corr. trib.,
30/2013, 2356.
341
On the state budget for financial year 2020.
342
Intervening on Legislative Decree No. 196 of 30th June 2003, paragraph
681 in fact added to:
- Article 2-sexies, second paragraph, regarding the processing of particular
categories of personal data necessary for reasons of important public interest, the one
performed by “persons who carry out duties in the public interest or connected with

188
with the tools to combat tax evasion343. The Budget Act of 2020 in fact
extended the possibility of the Revenue Office and Financial Police
using the information in the archive of financial relationships to make
an effective analysis of the risk of evasion, making use “also subject
to the pseudo-anonymisation of personal data”344 of the technologies,
processing and inter connections with other available databases in
order to identify useful risk criteria to get positions to emerge for
checking and to incentivise willing fulfilment” (paragraphs 682 and
686).

5. The acquisition of notices of “theoretical” crime.

Final reflections

Consultation of the archive of financial relationships and, more

generally, the databases available to the financial administration could

the exercise of public powers in the following subjects: […] prevention and combat
of tax evasion”;
- Article 2-undecies, among the scenarios for limiting the rights of the persons
concerned, situations where “an actual concrete harm could arise from the exercise of
those rights: [...] f-bis) to the protected interests on taxation and the performance of
activities to prevent and combat tax evasion”.
343
Article 23 of EU Regulation 2016/679 allows restrictions to the scope of
the obligations and rights provided for therein “when such a restriction respects the
essence of the fundamental rights and freedoms and is a necessary and proportionate
measure in a democratic society to safeguard: [...] public security; d) the prevention,
investigation, detection or prosecution of criminal offences or the execution of
criminal penalties, including the safeguarding against and the prevention of threats to
public security; e) an important economic or financial interest of the Union or of a
Member State, including monetary, budgetary and taxation a matters, public health
and social security; [...].
344
On this measure, in a critical sense, one should note the memorandum of
the Chairman of the Guarantee Authority for the protection of personal data on the
2020 budget bill, submitted to Commission 5a (Budget) of the Senate of the Republic
(12.11.2019): www.garanteprivacy.it.

189
make evidence of crime emerge through the exceeding of criminal
significance thresholds 345, also outside taxation346.
This eventuality is particularly significant for those in the
Financial Police, who cumulate the functions of economic/financial
police and judicial police, and poses the need to comply with Article
220 of the Code of Criminal Procedure under which “when evidence
of crime emerges during the inspection and oversight activities
provided for by laws or decrees, the acts necessary to ensure the
sources of evidence and collect anything else that may serve to apply
criminal law must be carried out in compliance with the provisions of
the code”.
In fact, if, for the United Sections, the pre-requisite for the
operating efficiency of Article 220 must be found in the “mere
possibility of attributing criminal significance in any case to the fact
that emerges from the administrative investigation and the time that it
emerges, regardless of the circumstance that it could refer to a certain
person”347, attention is not now directed towards the operators’ rules
of behaviour or the connected guarantee of non-usability set by the

345
According to the Supreme Court, during a tax audit, evidence of the crime
of false declaration must be considered to have emerged not on the evaded tax
threshold being exceeded, but even earlier, at the concrete probability of it being
exceeded. Thus, Cass., Sez. III pen., 3rd February 2015, No. 4919, in Riv. Guardia di
Finanza, 3/2015, 673, with note by T. RAFARACI, Reati tributari con soglia di
punibilità e applicazione dell’art. 220 disp. att. c.p.p.: la Cassazione rimarca i diritti
della difesa.
346
Not only the significant thresholds under Articles 4 (false declaration) and
5 (failure to declare) of Legislative Decree No. 74 of 10.03.2000, but also those in
Article 316-ter (Unlawful appropriation of funds to the detriment of the State) of the
Criminal Code, taking into account that the archive of financial relationships allows
the truthfulness of the single substitute declaration, concerning the notices needed to
determine the ISEE for the purposes of enjoying the facilitated social benefits, to be
checked.
347
Thus, Cass. sez. unite pen., 28th November 2001, n. 45477, in Il Fisco,
8/2002, 1178, with note by IZZO G., Le sezioni Unite limitano l’utilizzabilità di
dichiarazioni rese in sede ispettiva di vigilanza.

190
same Article 220348, for at least two types of reasons: on one hand,
because the information acquired from the tax register are significant
to the intents and purposes of Article 234 of the Code of Criminal
Procedure and, on the other, because the profile of greatest interest
concerns the possibility of also acquiring notices of “theoretical”
crime349, merely by cross-matching the data available in the
databases350.
The interoperability of the latter does not just concern the
public, domestic and European Union accounts revenue sector, but
also the expenditure. In safeguarding the proper use of them, a mention
should be given to the Sistema Informativo Antifrode [Anti-Fraud
Information System] (SIAF), a computer platform developed by the
Financial Police to enhance the combat against fraud at the expense of
the European Union budget: the information from the various

348
On this issue, see P. SORBELLO, La valutazione di sospetti, indizi e notizie
di reato nel passaggio (incerto) dalle attività ispettive alle funzioni di polizia
giudiziaria, in Riv. trim. dir. pen. cont., 2/2016, 125.
349
A check that the thresholds of criminal punishability for tax breaches have
been exceeded can also be performed “theoretically”. In fact, the Supreme Court
considered a conviction legitimate after the Financial Police determined the tax
evaded by a company be cross-matching the receivable and payable invoices in the
financial administration databases. If the contributor does not produce any
documentation aimed at “dismantling” the auditors’ calculations, the determination of
the evaded tax is enough to uphold the guilty verdict. Thus, Cass., sez. III pen., 21st
February 2018, n. 7871, in Leggi d’Italia.
350
One example concerns the failure to pay VAT as per Article 10-ter of
Legislative Decree 74/2000, “the value added tax based on the annual declaration,
within the term for paying the advance payment for the next tax period”, if the amount
exceeds the threshold of € 250,000 for each tax period. In the case in question, though
declaring its VAT debt, the contributor did not pay before the 27th December of the
next year, as provided for by Article 6, second paragraph, of Act No. 405 of 29th
December 1990: in order to ascertain the crime, at the end of the term, just identify
the VAT debts that are above the threshold of criminal significance and check their
non-payment via exclusive consultation of the tax register.

191
databases351 are processed automatically by the SIAF in view of
specific context and risk analyses.
Unlike traditional investigations based on knowledge of the
territory, observation and deduction, the new technologies make a
wide range of information resources available to the investigator to
direct the operational activity in a selective and much more profitable
manner.
The progressive expansion of the use of the financial
relationships archive, the one-to-one correspondence with the anti-
money-laundering discipline and, more generally, the interoperability
of the databases represent a significant crime policy lever because they
strengthen prevention of the crime, even if the right to the protection
of personal data is limited, with particular reference to the private area
of the contributors.
A similar worry must be filtered (or rather, re-dimensioned) by
rules and principles.
Firstly, under Article 1 of Directive 2016/680352, the right to the
protection of personal data concerns the protection of natural persons

351
On the basis of Article 3 of Legislative Decree 68/2001, regarding the inter-
institutional collaboration on economic and financial matters, the Financial Police has
over time signed agreements with foreign authorities and bodies with institutional
significance in the context of overseeing the public expenditure sector, also of
European origin. These agreements are based on the principle of inter-institutional
collaboration as per Article 3 of Legislative Decree 68/2001 or express provisions of
law which also affect the arrangement of the powers that can be used by the Corps
(e.g. the above-mentioned regulations on consultation of the financial relationships
archive). As regards the SIAF, these agreements include procedures for making
available information concerning the grantees and receivers of EU provisions for
autonomous info/operational development based on the economic/financial policing
prerogatives of the Financial Police. See Senate of the Republic, Lotta alle frodi in
danno delle uscite di bilancio dell’Unione europea [Fight against fraud at the expense
of the European Union budget]. Analysis Document No. 17, May 2018, 88 et seq.:
https://www.senato.it/service/PDF/PDFServer/BGT/01068773.pdf.
352
Article 1 of Legislative Decree 51/2018, which implements it, is the
corresponding article.

192
with regard to the processing carried out by the competent authorities
for the purposes of the prevention, investigation, detection and
prosecution of crimes: the circumstance that the experimental
procedure for the analysis of the risk of evasion 353 has so far only
concerned the tax position of partnerships and companies excludes the
question in a nutshell.
Secondly, also with reference to natural persons, “the right to
the protection of personal data is not an absolute right; it must be
considered in relation to its function in society and be balanced against
other fundamental rights, in accordance with the principle of
proportionality”354.
This balancing requirement, as can be seen in the preamble to
Directive 2016/680, also concerns police activities that “mainly
revolve around the prevention, investigation, detection and
prosecution of crimes, including police activities carried out without
prior knowledge of the criminal significance of a deed [and] for the
safeguard against and the prevention of threats [...] the fundamental
interests of a society protected by the law which can give rise to
crimes” (12), also including the relevant fiscal interest under Articles
2 and 53 of the Constitution. In pursuing these purposes, it is necessary
that “the competent authorities process the personal data collected for
the purposes of preventing, investigating, detecting or prosecuting
specific crimes outside such context to develop knowledge of criminal
activities and to connect the different crimes detected” (27): and this
is the risk analysis carried out thanks to the automated processing of
personal data. It is obvious that the limitation of the right to the
protection of personal data can be called lawful provided that the
processing is “ for specified, explicit and legitimate purposes [not]

353
See the provisions of the Director of the Revenue Office, No. 197357 of
31st August 2018 and No. 669173 of 8th August 2019.
354
Thus, Consideration No. 4 of Regulation 679/2016.

193
incompatible with the purposes of prevention, investigation, detection
and prosecution of crimes or the execution of criminal penalties” (29):
in fact, only in this case, can the processing be “for the execution of a
task carried out in the public interest” (35), in the absence of which
consultation of the databases would have criminal significance355.
The new technologies allow data to be analysed and compared,
returning a synthesis of complex information to direct the prevention
activities: the traditional search by investigators for notices of crime is
now also accompanied by the analysis of the metadata, i.e. the system
of relations between individually insignificant data. The automatic
cross-matching of the information in the databases can allow notices
of crime (e.g. tax ones) to be acquired almost immediately,
significantly affecting the number of unrecorded crimes without the
unlawful act being discovered when the prescription is already close.

355
In fact, it supplements “the offence provided for by Article 615-ter, second
paragraph, No. 1, of the Criminal Code: the conduct of a public official or person
tasked by a public service who, although being qualified and not breaching the formal
prescriptions imparted by the owner of a protected IT or ICT system to limit access to
it, accesses or stays in the system for ontologically extraneous reasons or ones that are
in any case different to those for which he/she is only granted the right of access”.
Thus, Cass. sez. un. pen., 8th September 2017, No. 41210, in Dir. pen. proc., 4/2018,
506, with note by R. FLOR, La condotta del pubblico ufficiale fra violazione della
voluntas domini, “abuso” dei profili autorizzativi e “sviamento di potere”.

194
 BIBLIOGRAPHY

Various Authors, Manual de Introduccion al Derecho Penal, dir. J.

Lascuraín Sánchez, Agencía Estatal BOE, 2019
Various Authors, Predictive Policing. The Role of Crime Forecasting in
Law Enforcement Operations, in RAND. Safety and Justice Program, 2013, 19 et seq.,
available at the following address:
https://www.rand.org/content/dam/rand/pubs/research_reports/RR200/RR233/RAN
D_RR233.pdf
Various Authors, Procedure penali d’Europa, edited by M. Delmas-Marty,
M. Chiavario, Cedam, 2001
ALETRAS N., TSARAPATSANIS D., PREOTIUC-PIETRO D., LAMPOS V.,
Predicting judicial decisions of the European Court of Human Rights: a Natural
Language Processing perspective, in PeerJ Computer Science, 2016
AMIGONI F., SCHIAFFONATI V., SOMALVICO M., in Enciclopedia della
Scienza e della Tecnica, 2008, in www.treccani.it
AMODIO E., Perizia e consulenza tecnica nel quadro probatorio del nuovo
processo penale, in Cass. pen., 1989, 170 et seq.
ANGWIN J., LARSON J., MATTU S., KIRCHNER L., How We Analyzed the
COMPAS Recidivism Algorithm in www.propublica.org, 23rd May 2016
ANGWIN J., LARSON J., MATTU S., KIRCHNER L., Machine Bias: there’s
software used across the country to predict future criminals. And it’s biased against
blacks, in www.propublica.org, 23rd May 2016
ARMSTRONG T., Managing for 21st Century Crime Prevention in Memphis,
in www.managementexchange.com
BARBERIS M., Cosa resta del sillogismo giudiziale? Riflessioni a partire da
Beccaria, in Materiali per una storia della cultura giuridica, 1/2015
BARRY -JESTER A.M., CASSELMAN B., GOLDSTEIN D., Should Prison
Sentences Be Based on Crimes That Haven’t Been Committed Yet?, in
fivethirtyeight.com, 4th August 2015
BASHIR A., Pennsylvania’s Misguided Sentencing Risk-Assessment Reform,
in The Regulatory Review, 5th November 2019
BASILE F., Intelligenza artificiale e diritto penale: quattro possibili percorsi
di indagine, in Dir. pen. uomo, 29th September 2019
BEN-HARI D., FRISCH Y., LAZOVKI A., ELDAN U., GREENBAUM D., Artificial
Intelligence in the Practice of Law: An Analysis and Proof of Concept Experiment, in
Richmond Journal of Law and Technology, 2/2017
BENESTY M., The Judge Statistical Data Ban. My Story, in
www.artificiallawyer.com, 7th June 2019.
BIELLI D., Periti e consulenti nel nuovo processo penale, in Giust. pen.,
2/1991, 65 et seq.
BIG BROTHER WATCH TEAM, A closer look at Experian Big Data and
Artificial Intelligence in Durham Police, in bigbrotherwatch.org.uk, 6th April 2018

195
 N. BOLDRINI, Deep Learning, cos’è l’apprendimento profondo, come
funziona e quali sono i casi di applicazione, in www.ai4business.it, 9th August 2019
BONFANTI A., Big data e polizia predittiva: riflessioni in tema di protezione
del diritto alla privacy e dei dati personali, in Rivista di diritto dei media, 3/2018
BORSARI R., Intelligenza Artificiale e responsabilità penale: prime
considerazioni, in www.medialaws.eu, 3/2019
BORSARI R., Intelligenza Artificiale e responsabilità penale: prime
considerazioni, in MediaLaws, 3/2019
BRANTINGHAM P.J. – BRANTINGHAM P.L., Environmental criminology: from
theory to urban planning practice, in Studies on Crime and Crime Prevention, 7/1998
BRATING K., LESTER J.C., CALLAWAY C.B., Automating Judicial Document
Drafting: A Discourse-Based Approach, in Artificial Intelligence and Law, 1998
BRKAN M., Do algoritms Rule the World? Algorithmic Decision-Making in
the Framework of the GDPR and Beyond, in Electronic Journal, 2017, 1 et seq.
BURCHARD C., L’intelligenza artificiale come fine del diritto penale? Sulla
trasformazione algoritmica della società, in Riv. it. dir. proc. pen., 4/2019, 1908 et
seq.
CALAMANDREI P., Processo e democrazia, in Opere giuridiche, I, Roma
TrE-Press, 2019
CALLARI F., Il periculum libertatis costituito dal rischio di realizzazione di
determinati reati e le misure cautelari: il fine giustifica i mezzi?, in Dir. pen. cont.,
12th November 2012
CAPLAN J.M. – KENNEDY L.W. (editor), Risk Terrain Modeling Manual:
Theoretical Framework and Technical Steps of Spatial Risk Assessment for Crime
Analysis, CreateSpace Independent Publishing Platform, 2010
CAPPELLINI A., Machina delinquere non potest? Brevi appunti su
intelligenza artificiale e responsabilità penale, in Discrimen, 27th March 2019
CARBONE M., Archivio dei rapporti finanziari: esigenze di
razionalizzazione tra normativa fiscale e antiriciclaggio, Corr. trib., 30/2013, 2356
CARCENI L., Processo penale minorile, in Enc. Dir., Agg., vol. IV, Giuffrè,
2000, 1018 et seq.
CARLO S., Big Brother Watch’s written evidence on algorithms in the justice
system for the Law Society’s Technology and the Law Policy Commission, in
bigbrotherwatch.org.uk
CARRER S., Se l’amicus curiae è un algoritmo: il chiacchierato caso Loomis
alla Corte Suprema del Wisconsin, in Giurisprudenza Penale Web, 4/2019
CASONATO C., Intelligenza artificiale e giustizia: potenzialità e rischi,
DPCE – online, 3/2020
CASTELLI C., PIANA D., Giustizia predittiva. La qualità della giustizia in due
tempi, in Quest. giust., 4/2018, 153 et seq.
CEVOLANI G. – CRUPI V., Come ragionano i giudici: razionalità, euristiche
e illusioni cognitive, in Criminalia, 2017, 181 et seq.
CHAINEY S. – TOMPSON L., Engagement, Empowerment and Transparency:
Publishing Crime Statistics using Online Crime Mapping, in Policing, 2012

196
 CHIEL E., Secret Algorithms That Predict Future Criminals Get a Thumbs
Up From Wisconsin Supreme Court, in www.splinternews.com, 27th July 2016
CONTISSA G., LASAGNI G., SARTOR G., Quando a decidere in materia penale
sono (anche) algoritmi e IA: alla ricerca di un rimedio effettivo, in Diritto di internet,
4/19
CORMEN T.H., LEISERSON C.E., RIVEST R.L., STEIN C., Introduction to
Algorithms, 3rd ed., MIT Press, 2009
CRAGLIA M. (Ed.), ANNONI A., BENCZUR P., BERTOLDI P., DELIPETREV P.,
DE PRATO G., FEIJOO C., FERNADEZ M ACIAS E., GOMEZ E., IGLESIAS M., JUNKLEWITZ
H., LÓPEZ COBO M., MARTENS B., NASCIMENTO S., NATIVI S., POLVORA A., SANCHEZ
I., TOLAN S., TUOMI I., VESNIC ALUJEVIC L., Artificial Intelligence – A European
Perspective, EUR 29425 EN, Publications Office, 2018
D’AGOSTINO L., Gli algoritmi predittivi per la commisurazione della pena.
A proposito dell’esperienza statunitense nel c.d. evidence-based sentencing, in Dir.
pen. cont., 2/2019, 354 et seq.
DAVIS J.P., Law without mind. AI, Ethics and Jurisprudence, in University
of San Francisco Law research Paper, 2018
DE LUCA C., Gli accertamenti sulla personalità dell’autore del reato
minorenne e il divieto di perizia psicologica nel rito ordinario: riflessioni e nuove
prospettive, in Cass. pen., 2018, 2146 et seq.
DI GIOVINE O., Il judge-bot e le sequenze giuridiche in materia penale
(intelligenza artificiale e stabilizzazione giurisprudenziale), in Cass. pen., 3/2020, 951
DI NICOLA A. – ESPA G. – BRESSAN S. – DICKSON M.M. – A. NICOLAMARINO,
Metodi statistici per la predizione della criminalità. Rassegna della letteratura su
predictive policing e moduli di data mining, in eCrime Working Papers, 2/2014, 5 et
seq.
DIENER A., Pennsylvania’s Proposed, Questionably Constitutional, Risk
Assessment Instrument, in Harvard Civil Rights – Civil Liberties Law Review, 17th
October 2019
DIJKSTRA, E.W. The Threats to Computing Science, 1984, available at
www.cs.utexas.edu
DONINI M., La riserva di codice (art. 3-bis c.p.) tra democrazia normante e
principi costituzionali. Apertura di un dibattito, in Leg. pen., 20th November 2018
DRESSEL J., FARID H., The accuracy, fairness, and limits of predicting
recidivism, in Science Advances, 4/2018
ECK J. – CHAINEY S. – CAMERON J. – LEITNER M. – WILSON R., Mapping
Crime: Understanding Hot Spots, in National Institute of Justice, 2005, available at
the following address: https://nij.ojp.gov/library/publications/mapping-crime-
understanding-hot-spots
ENGLICH B. – MUSSWEILER T. – STRACK F., Playing Dice With Criminal
Sentences: The Influence of Irrelevant Anchors on Experts’ Judicial Decision Making,
in Personality and Social Psychology Bulletin, 2/2006, 194

197
 EPIC (Electronic Privacy Information Center), Algorithms in the Criminal
Justice System: Risk Assessment Tools, available at the following link:
https://epic.org/algorithmic-transparency/crim-justice/
ERAMO F., Il divieto di perizie psicologiche nel processo penale: una nuova
conferma dalla Cassazione, in Dir. pen. proc., 7/2007, 931 et seq.
ETXEBARRIA ZARRABEITIA X., Medidas de seguridad: presupuestos de su
aplicación, in Cuadernos penales José María Lindón, 10/2014, 125 et seq.
FALSITTA G., Epicedio per il segreto bancario nei confronti del Fisco, in
Riv. dir. trib., 2/1992, 566
FIANDACA G., Pena “patteggiata” e principio rieducativo: un arduo
compromesso tra logica di parte e controllo giudiziale, in Foro it., 1990, I, 2385
FLOR R., La condotta del pubblico ufficiale fra violazione della voluntas
domini, “abuso” dei profili autorizzativi e “sviamento di potere”, in Dir. pen. proc.,
4/2018, 506
FORGIONE I., Il caso dell’accesso al software MIUR per l’assegnazione dei
docenti – T.A.R. Lazio Sez. III bis, 14th February 2017, No. 3769, in Giornale di
diritto amministrativo, 2018, 647 et seq.
FREEMAN K., Algorithmic Injustice: How the Wisconsin Supreme Court
Failed to Protect Due Process Rights in State v. Loomis, in N.C. J.L. & Tech., 12/2016
FRIEND Z., Predictive Policing: Using Technology to Reduce Crime, in FBI
Law Enforcement Bulletin, 2013
GABELLINI E., La “comodità nel giudicare”: la decisione robotica, in Riv.
trim. dir. proc. civ., 4/2019, 1305 et seq.
GALANTINI N., Il principio di obbligatorietà dell’azione penale tra interesse
alla persecuzione penale e interesse all’efficienza giudiziaria, in Dir. pen. cont.,
23.9.2019.
GALGANI B., Considerazioni sui “precedenti” dell’imputato e del giudice al
cospetto dell’IA nel processo penale, in Sist. pen., 4/2020, 87 et seq.
GARAPON A., LASSÈGUE J., Justice digitale. Révolution graphique et rupture
anthropologique, Puf, 2018.
GARBOLINO P., Nuovi strumenti logici e informatici per il ragionamento
giudiziario: le reti bayesiane, in Cass. pen., 1/2007, 326 et seq.
GIALUZ M., Quando la giustizia penale incontra l’intelligenza artificiale:
luci e ombre dei risk assessment tools tra Stati Uniti ed Europa, in Dir. pen. cont.,
29th May 2019, 1 et seq.
GIALUZ M., Quando la giustizia penale incontra l’intelligenza artificiale:
luci e ombre dei Risk Assessment Tools tra Stati Uniti ed Europa, in
www.penalecontemporaneo.it, 28th May 2019, 21
GIANNINI I., Il dibattito sull’ammissibilità della perizia e della consulenza
criminologica nel processo penale, in Rass. penit. crim., 3/2003, 87 et seq.
GILLESPIE T., The relevance of Algorithms, in GILLESPIE T., BOCZKOWSKI P.,
FOOT K., Media Technologies, MIT Press, 2014, 1 et seq.
GIUPPONI T., Sicurezza personale, sicurezza collettiva e misure di
prevenzione. La tutela dei diritti fondamentali e l’attività di intelligence. Intervention

198
at the “Collective Security and Fundamental Rights” seminar held at the University of
Ferrara on 26th September 2007, www.forumcostituzionale.it.
GONZÁLEZ ÁLVAREZ J.L. – SANTOS HERMOSO J. – CAMACHO COLLADOS M.,
Policía predictiva en España. Aplicación y retos futuros, in Behavior & Law Journal,
6/2020, 26 et seq.
GONZÁLEZ J.L. – GARRIDO M.J. – LÓPEZ OSSORIO J.J. – MUÑOZ J.M. –
ARRIBAS A. – CARBAJOSA P. – BALLANO E., Revisión pormenorizada de homicidios de
mujeres en las relaciones de pareja en España, in Anuario de Psicología Jurídica,
2018, 28 et seq.
GONZÁLEZ J.L. – MUÑOZ V. – CALCERRADA M.L. – SOTOCA A., Perfil
psicosocial del incendiario forestal español privado de libertad, in Behavior & Law
Journal, 3/2017, 26 et seq.
HANNEM S., SANDERS C. B., SCHNEIDER C. J., DOYLE A., CHRISTENSEN T.,
Security and Risk Technologies in Criminal Justice: Critical Perspectives, Canadian
Scholars, 2019
HART N. – YOHANNES M., Evidence Works: Cases Where Evidence
Meaningfully Informed Policy, in Bipartisan Policy Center, 2019, available in full at
the following address: https://bipartisanpolicy.org/report/evidenceworks/
HESTER R., Risk Assessment at Sentencing. The Pennsylvania Experience,
in VARIOUS AUTHORS, Predictive Sentencing: Normative and Empirical Perspectives,
edited by J.W. de Keijser, J.V. Roberts, J. Ryberg, Bloomsbury Publishing, 2019, 213
et seq.
HUQ A.Z., Racial Equity in Algorithmic Criminal Justice, in Duke Law
Journal, 2019, 1043 et seq.
ITZCOVICH G., Il diritto come macchina. Razionalizzazione del diritto e
forma giuridica in Max Weber, in Materiali per una storia della cultura giuridica, 31/2,
2001, 365
ITZCOVICH G., Sulla metafora del diritto come macchina, in Dir. quest.
pubbl., 9/2009, 379
KAPLAN J., Intelligenza artificiale. Guida al futuro prossimo, Luiss
University Press, 2017
KAPLAN J., Le persone non servono. Lavoro e ricchezza nell’epoca
dell’intelligenza artificiale, translated by Ilaria Veronica Tomasello, Luiss University
Press, 2016.
KEHL D., GUO P., KESSLER S., Algorithms in the Criminal Justice System:
Assessing the Use of Risk Assessments in Sentencing (Responsive Communities
Initiative, Berkman Klein Center for Internet & Society, Harvard Law School), July
2017, available at the following link:
https://cyber.harvard.edu/publications/2017/07/Algorithms
KEHL D., GUO P., KESSLER S., Algorithms in the Criminal Justice System:
Assessing the Use of Risk Assessments in Sentencing (Responsive Communities
Initiative, Berkman Klein Center for Internet & Society, Harvard Law School), 7/2017

199
 KING T., AGGARWAL M., TADDEO M., FLORIDI L., Artificial Intelligence
Crime: An Interdisciplinary Analysis of Foreseeable Threats and Solutions, in Science
and Engineering Ethics, 2019
KOEPKE J.L., ROBINSON D.G, Danger Ahead: Risk Assessment and the
Future of Bail Reform, in Washington Law Review, 4/2018, 1725 et seq.
KREMER J., The End of Freedom in Public Places? Privacy problems arising
from surveillance of the European public space, University of Helsinki, 2017,
available at helda.helsinki.fi.
LANGFORD M., MADSEN M.R., France Criminalises Research on Judges, in
verfassungsblog.de, 22nd June 2019
LÓPEZ OSSORIO J.J. – GONZÁLEZ ÁLVAREZ J.L. – MUÑOZ VICENTE J.M. –
URRUELA C. – ANDRÉS PUEYO A., Validation and Calibration of the Spanish Police
Intimate Partner Violence Risk Assessment System (VioGén), in Journal of Police and
Criminal Psychology, 2019
LÓPEZ OSSORIO J.J. – GONZÁLEZ J.L. – BUQUERÍN S. – GARCÍA L.F. – BUELA
CASAL G., Risk Factors related to intimate partner violence police recidivism in
Spain, in International Journal of Clinical and Health Psychology, 2017, 107 et seq.
LUM K., ISAAC W., To predict and serve?, in Significance, vol. 13, 10/2016,
14-19
LUM K., Predictive Policing Reinforces Police Bias, in Human Rights Data
Analysis Group, available at www.hrdag.org
LUPÀRIA L., Prova giudiziaria e ragionamento artificiale: alcune possibili
chiavi di lettura, in Il concetto di prova alla luce dell’intelligenza artificiale, edited
by J. Sallantin, J.J. Szczeciniarz, Giuffrè, 2005
MALDONATO L., Algoritmi predittivi e discrezionalità del giudice: una
nuova sfida per la giustizia penale, in Dir. pen. cont., 2/2019, 401 et seq.
MALGIERI G., COMANDÈ G., Why a Right to Legibility of Automated
Decision-Making Exists in the General Data Protection Regulation, in International
Data Privacy Law, 2017, vol. 7, 243 et seq.
MANES V., L’oracolo algoritmico e la giustizia penale: al bivio tra
tecnologia e tecnocrazia, in Intelligenza artificiale. Il diritto, i diritti, l’etica, edited
by U. Ruffolo, Giuffrè, 2020, 547
MANTOVANI F., La “perenne crisi” e “la perenne vitalità” della pena. E la
“crisi di solitudine” del diritto penale, in DOLCINI E. - PALIERO C.E. (edited by), Studi
in onore di Giorgio Marinucci, I, Giuffrè, 2006, 1173
MARTINI A., Essere pericolosi. Giudizi soggettivi e misure personali,
Giappichelli, 2017
MARTUCCI P., Il contributo del criminologo nel processo penale: un
problema ancora aperto, in Dir. pen. proc., 6/2004, 744 et seq.
MASSARO A., Determinatezza della norma penale e calcolabilità giuridica,
Editoriale Scientifica, 2020
MASSARO A., I reati “contro gli animali” tra aspirazioni zoocentriche e
ineliminabili residui antropocentrici, in Cult. dir., 1-2/2018

200
 MASTROBUONI G. (Lead Academic) Impact: Imagine being able to predict a
crime in the future, Research case study by University of Essex, in www.essex.ac.uk
MASTROBUONI G., Crime is Terribly Revealing: Information Technology
and Police Productivity, in The Review of Economic Studies, vol. 87, 11/ 2020, 2732
MC CARTHY J., MINSKY M.L., ROCHESTER N., SHANNON C.E., A Proposal
for the Dartmouth Summer Research Project on Artificial Intelligence, 1955, available
at www-formal.stanford.edu
MCCARTHY J., MINSKY M.L., ROCHESTER N., SHANNON C.E., A proposal for
the Dartmouth summer research project on Artificial Intelligence, 31st August 1955,
in www-formal.stanford.edu
MORABITO C., La chiave del crimine, in www.poliziadistato.it, July 2015
C. MORELLI, Giustizia predittiva: il progetto (concreto) della Corte
d’appello di Brescia, in www.altalex.it, 8th April 2019.
MORO P., Intelligenza artificiale e professioni legali. La questione del
metodo, in Journal of Ethics and Legal Technologies, 2019
MOSCARINI P., La perizia psicologica e il ‘giusto processo’, in Dir. pen
proc., 8/2006, 929 et seq.
NIEVA-FENOLL J., Intelligenza artificiale e processo, translated by P.
Comoglio, Giappichelli, 2019
NISSAN E., Digital Technologies and Artificial Intelligence’s Present and
Foreseeable Impact on Lawyering, Judging, Policing and Law Enforcement, in
AI&Soc., 2017
NUVOLONE P., Appunti e spunti tra precetti e sanzioni, in Riv. it. dir. pen.,
1956, 441
O’DONNEL R.M., Challenging Racist Predictive Policing Algorithms under
the Equal Protection Clause, in New York University Law Review, 2019
OCCHIUZZI B., Algoritmi predittivi: alcune premesse metodologiche, in Dir.
pen. cont., 2/2019, 391 et seq.
OSWALD M., GRACE J., URWIN S., BARNES G.C., Algorithmic risk assessment
policing models: lessons from the Durham HART model and “Experimental”
proportionality, in Information and Communications Technology Law, 2/2018, 223 et
seq.
PALIERO C.E ., L’economia della pena (a work in progress), in Riv. it. dir.
proc. pen., 4/2005, 1396
PARK J., Your Honor, AI, in Harvard International Review, 3rd April 2020.
PARODI C., SELLAROLI V., Sistema penale e intelligenza artificiale: molte
speranze e qualche equivoco, in Dir. pen. cont., 6/2019
PELISSERO M., Pericolosità sociale e doppio binario. Vecchi e nuovi modelli
di incapacitazione, Giappichelli, 2008
PELLICCIA R., Polizia predittiva: il futuro della prevenzione criminale?, in
www.cyberlaws.it, 9th May 2019
POGGI F., Il diritto meccanico. La metafora del diritto come macchina e i
suoi limiti, in Dir. quest. pubbl., 9/2009, 395

201
 PROVOST F. – FAWCETT T., Data Science for Business: What you need to
know about data mining and data-analytic thinking, O’Reilly Media, 2013
QUATTROCOLO S., Equo processo penale e sfide della società algoritmica,
in Riv. BioDiritto, 1/2019, 135 et seq.
QUATTROCOLO S., Intelligenza artificiale e giustizia: nella cornice della
Carta etica europea, gli spunti per un’urgente discussione tra scienze penali e
informatiche, in Leg. pen., 18th December 2018
QUATTROCOLO S., Questioni nuove e soluzioni antiche? Consolidati
paradigmi normativi vs rischi e paure della giustizia digitale “predittiva”, in Cass.
pen., 4/2019, 1478 et seq.
RAFARACI T., Reati tributari con soglia di punibilità e applicazione dell’art.
220 disp. att. c.p.p.: la Cassazione rimarca i diritti della difesa. in Riv. Guardia di
Finanza, 3/2015, 673.
RATCLIFFE J.H., Intelligence-Led Policing, Routledge, 2016;
REYLING D., Beyond court digitalization with ODR, in International Journal
for Court Administration, 8(2), 1 et seq.
RICOZZI C. –DI PAOLO N., La funzione di polizia economica e finanziaria,
in Riv. Guardia di Finanza, 5/2003, 1485
RIVELLO P., voce Perito e perizia, Dig. disc. pen., IX, 474 et seq.
RONSIN X., LAMPOS V., MAIT̂ REPIERRE A., Questioni specifiche della
giustizia penale: prevenzione del reato, del rischio di recidiva e valutazione del livello
di pericolosità, Appendice I: Studio approfondito sull’utilizzo dell’intelligenza
artificiale [IA] nei sistemi giudiziari, segnatamente delle applicazioni
dell’intelligenza artificiale al trattamento di decisioni e dati giudiziari, in www.coe.int
RONSIN X., LAMPOS V., MAÎTREPIERRE V., In-depth study on the use of AI
in judicial systems, notably AI applications processing judicial decisions and data, in
Appendix I, European ethical Charter on the use of Artificial Intelligence in judicial
systems and their environment, in www.coe.int.
RULLI E., Giustizia predittiva, intelligenza artificiale e modelli
probabilistici. Chi ha paura degli algoritmi?, in Analisi giuridica dell’economia,
2/2018
RUSSEL S., NORVIG P., Intelligenza artificiale. Un approccio moderno, I, II
ed., Pearson 2005
SAMOILI S., LÓPEZ COBO M., GÓMEZ E., DE PRATO G., MARTÍNEZ-PLUMED
F., DELIPETREV B., AI Watch. Defining Artificial Intelligence. Towards an operational
definition and taxonomy of artificial intelligence, EUR 30117 EN, Publications Office
of the European Union, 2020
SAPONARO A., L’esame della personalità del reo nel processo penale:
evoluzione, involuzione, modelli alternativi, prospettive, Cacucci, 2000
SCHWAB K., La quarta rivoluzione industriale, Franco Angeli, 2016
SCURICH N., The case against categorical risk estimates, in Behavioral
Science Law, 2018, 1 et seq.
SEARLE J.R., Minds, brains and programs, in Behavioral and Brain
Sciences, 1980

202
 SIGNORATO S., Giustizia penale e intelligenza artificiale. Considerazioni in
tema di algoritmo predittivo, in Riv. dir. proc., 2020, 614 et seq.
SIMONCINI A., L’algoritmo incostituzionale: intelligenza artificiale e il
futuro delle libertà, BioLaw Journal – Rivista di BioDiritto, 1/2019
SORBELLO P., La valutazione di sospetti, indizi e notizie di reato nel
passaggio (incerto) dalle attività ispettive alle funzioni di polizia giudiziaria, in Riv.
trim. dir. pen. cont., 2/2016, 125
SORBELLO P., Politica criminale ed osservanza delle regole. Riflessioni su
limiti e possibilità di conversione al razionale dei comportamenti, in Riv. it. dir. proc.
pen., 4/2017, 1934
SORO A., Proteggere i dati per governare la complessità. Speech of the
Chairman of the Guarantee Authority for the protection of personal data, Rome, 2017
SULLIVAN T.J. – PERRY W.L., Identifying Indicators of Chemical,
Biological, Radiological, and Nuclear (CBRN) Weapons Development Activity in
Sub-National Terrorist Groups, in Journal of the Operational Research Society, 2004
SZWARC C., La justice predictive: une autre justice?, in Le Mag des Avocats,
34/2017, 5 et seq., available at the following link:
https://www.anased.fr/publications/mag34/files/assets/basic-html/page-5.html#
TATA C., The Application of Judicial Intelligence and ‘Rules’ to Systems
Supporting Discretionary Judicial Decision-Making, in Artificial Intelligence and
Law, 1998
TESAURO F., Istituzioni di diritto tributario, Giappichelli, 2011
TONDI V., Il Bail. La libertà su cauzione negli ordinamenti anglosassoni,
Cedam-Wolters Kluwer, 2016
TRANCHINA G., Il divieto di perizia psicologica sull’imputato: una
limitazione sicuramente anticostituzionale, in Riv. it. dir. proc. pen., 1971, 1325 et
seq.
TRAVERSI A., Intelligenza artificiale applicata alla giustizia: ci sarà un
giudice robot?, in Quest. giust., 10th April 2019.
TRAVERSI A., Intelligenza artificiale applicata alla giustizia: ci sarà un
giudice robot?, in Quest. giust., 10th April 2019.
TREVISI C., La regolamentazione in materia di Intelligenza artificiale,
robot, automazione: a che punto siamo, in Medialaws, 2018, 1 et seq.
TRIDICO B.D., L’utilizzo dell’anagrafe dei rapporti finanziari ai fini
dell’attività di controllo fiscale, Relazione Corte dei Conti – Sez. centrale di controllo
sulla gestione delle Amministrazioni dello Stato (26th July 2017, n. 11/2017/G)
TURING A.M., Computing machinery and intelligence, in Mind, LIX, 236,
10/1950
TUZET G., Il diritto non è una macchina, in Dir. quest. pubbl., 401 et seq.
TUZET G., L’algoritmo come pastore del giudice? Diritto, tecnologie, prova
scientifica, in MediaLaws, 2019
TUZET G., Sul possibile moto della macchina, in Dir. quest. pubbl., 9/2009,
359

203
 UBERTIS G., Intelligenza artificiale, giustizia penale, controllo umano
significativo, in Sist. pen., 11th November 2020, 1 et seq.
VELLUZZI V., La preziosa ingenuità: Beccaria, lo spirito della legge e il
sillogismo giudiziale, in Riv. int. fil. dir., 2014
VENTURI M., KeyCrime© – La chiave del crimine, in PrimoPiano, 12/2014,
4, available at www.onap-profiling.org
VEZZADINI S., Profilo geografico e crime mapping. Il contributo della
criminologia ambientale allo studio del delitto, in Scena del crimine e profili
investigativi. Quale tutela per le vittime?, edited by R. Bisi, Franco Angeli, 2006
VINCENT G.M., CHAPMAN J., COOK N.E., Risk-Needs Assessment in Juvenile
Justice: Predictive Validity of the SAVRY, Racial Differences, and the Contribution of
Needs Factors, in Criminal Justice & Behavior, 2011, 47 et seq.
VIOLA L., L’intelligenza artificiale nel procedimento e nel processo
amministrativo: lo stato dell’arte, in Foro amm., 2018, 1598 et seq.
VLAHOS J., Come anticipare il crimine, in Le scienze, 2012
VOGLER R., FOULADVAND S., Standards for making factual determinations
in arrest and pre-trial detention: a comparative analysis of law and practice, in
VARIOUS AUTHORS, Comparative Criminal Procedure, edited by J. E. Ross, S.C.
Thaman, Edward Elgar Publishing, 2016
WEBER M., Economia e società, IV, Edizioni di Comunità, 1995
WIDGERY A., The Statutory Framework of Pretrial Release, in
www.ncsl.org, 8th November 2020

204