Public perception of algorithms in

society: accounting for the algorithmic

public opinion . The Algorithmic Public Opinion:
Algocount 2020-2022
a Policy Overview

founded by : Urbano Reviglio, Università degli Studi di Milano

The algocount project seeks to contribute to concerns on the

political and policy implications of the rise of algorithms. The diffusion
of an “algorithmic public opinion” specifically regards the concern that
algorithmic systems disrupt previous patterns of individual and public
opinion formation. Few analyses explored this phenomenon throughout
a multidisciplinary comprehensive policy lens. The algocount project
seeks to address this gap by discussing innovative policy approaches
and proposals to the challenges of the current algorithmic public
opinion. Firstly, I introduce the main algorithms, concerns and actors
involved in the algorithmic public opinion. Then, I discuss the main
policy approaches to the systemic challenges of the most relevant
algorithmic systems: recommender systems, disinformation,
political microtargeting, content moderation and social bots. The
goal of this report is to provide a critical, comprehensive and long-
term oriented policy overview of the algorithmic public opinion.

Table of Contents

10

18

29
52
Executive Summary
→ The algorithmic public opinion comprises several algorithms
which profoundly influence worldwide information flows -
information production, distribution and consumption - and,
therefore, intermediating – not determining – opinion formation
both at the individual and collective level.
→ The most influencing algorithms in this context are arguably
personalization algorithms that act as fundamental global
gatekeepers. Yet, other relevant algorithms are the ones employed
in content moderation (to detect and remove content), social bots
(to manipulate content diffusion), and, among others, those who
help to create and edit content.
→ More than decreased media diversity (e.g. filter bubble theory and
echo chambers), the most concerning issues in the algorithmic
public opinion are algorithmic censorship, algorithmic
manipulation and, more generally, disinformation.
→ It is paramount to allow researchers to access wider datasets in
a legal and transparent manner as well as to let them run on-
platform experiments capable of detecting more reliable causal
relationships. This is a condition sine qua non for collecting
evidence for effective policies. Importantly, policies for the
algorithmic public opinion can even increase the quality of public
policies more generally.
→ The question of regulating the algorithmic public opinion is
part of a complex regulatory system involving many different
legislative acts and normative approaches. Policy-making still
lacks a shared vocabulary or frameworks for approaching the
incoming challenges and new models for digital governance will
likely need to be developed. The algorithmic public opinion is in
fact particularly challenging to policy-makers as it intersects with
a transformation of media itself, due to media convergence, media
globalization and the rise of global digital platforms.
→ It is desirable to foster citizens’ active participation in the design
and governance of digital platforms, rather than relying on
top-down regulations by a technocratic elite. Similarly, a truly
multistakeholder governance has to allow civil society a voice,
“loud” enough to be heard. A truly multistakeholder governance
approach in which civil society has finally a voice is necessary.
→ The role of ethics in mediating the regulatory process is essential.
Despite widespread phenomena such as ethics washing and ethics
bashing who threaten to limit its potential relevance and influence,
ethical debates are necessary to develop better and more inclusive
policies and guidelines and, therefore, deserve more attention
from all the involved stakeholders.
→ As traditional news media is still struggling with the opportunities
and challenges of digital societies as well as the emerging
algorithmic public opinion, it is fundamental to understand its
essential role in re-mediating the algorithmic public opinion. News
quality, thus a well-informed public opinion, highly depends on
 the ability of news media to remain politically and economically
independent so as to provide high quality journalism.
→ US and European policymakers have mostly focused on the latest
controversies and the biggest players. To tackle urgent policies
we need to understand algorithms as expansive socio-technical
phenomena that functions in many contexts and takes many
forms. Expanding the scope and range of research is critical to
developing sound policy. The range of contentious phenomena
cannot be captured with sectorial approaches. Fundamentally, the
largest, US-based platforms do not provide a reliable guide for the
entire social media ecology.
→ For a sustainable algorithmic public opinion, it is needed a
paradigm change. Admittedly, the main concerns and risks
associated with the algorithmic public opinion are mainly rooted
in the perverse dynamics arising from the current mainstream
business model based on advertising. This model is at the heart of
surveillance capitalism which is widely believed it needs structural
reforms.
→ It is necessary to focus on policies to reinforce the quality and
effectiveness of policies themselves. There are indeed structural
concerns of lobbying pressures from a variety of actors, in
primis big tech companies, to consistently undermine the most
innovative and effective impetuses of new regulations. Similarly,
there should be stricter rules on the phenomenon of revolving
doors.
→ Finally, we should avoid naive expectations of the Internet as an
ideal public sphere: there are always new challenges, and human
and machines mistakes will inevitably raise further concerns.
It is a never-ending challenge the governance of the algorithmic
public opinion.
 An Introduction to the Algorithmic
Public Opinion

Algorithms are everywhere nowadays. From traffic lights to election

machines, from stock markets to pacemakers: there is a wide range
of algorithmic variability, complexity and relevance. In the media
landscape, they are used in various contexts for different purposes; from
the selection of personalized content to the detection and removal of
fake news and harmful or copyrighted content, from the selection of
news for journalists to the automated creation of news articles, from the
usage of social bots to the creation of deepfakes. Algorithms are indeed
becoming an essential intermediary of how information is produced,
disseminated and consumed, and thus how public opinion is formed,
giving rise to what we define as an “algorithmic public opinion” (cfr. D6).

What are algorithms and how they really work, however, is not
easy to explain, if at all. In public discussions, the term “algorithms”
has indeed become a ‘sloppy shorthands, slang terms for the act of
mistaking multipart complex systems for simple, singular ones’ (Bogost,
2015). Algorithms are more than one (neutral) technology. As Lessig
(1999) famously pointed out, rules are also implicit in code – in all
the algorithms that select, prioritize, remove and, ultimately, affect
the public opinion. More broadly, an algorithm can be intended as “a
socio-economic construct, that is, a technology that is embedded in
organisations with their own goals, values and fundamental freedoms,
and that mediate and impact interactions with the human/economic/
social environment in which they are functioning” (Helberger et al,
2019, p. 3). There are, indeed, several human influences embedded
into algorithms, such as criteria choices, training data, semantics,
and interpretation. Algorithms need to be seen and analyzed as a
complex assemblage of procedures, individuals and teams, ideology,
datasets constantly changing and difficult to reduce to a singular,
simple concept. To fully understand them and their consequences is
therefore fundamental to understand both their technical complexity
and variability on the one hand, and their fundamentally social
embeddedness on the other.

Types of Algorithmic Systems

There are several algorithmic systems which influence the formation

of public opinion. These can affect the news ecosystem on at least three
levels: news production and distribution, individual users, the broader
media landscape and society more generally. Algorithms exert power
through decisions they make, for example, in prioritizing information
(e.g. search engines), filtering news (e.g. Facebook, Youtube etc.),
making classifications (e.g. Airbnb through the reputation system) or
finding associations (e.g. Google Flu Trends). More generally, algorithms
can automate several actions; from smart tools that assist journalists
in producing their stories and fully automated production of news
stories (so-called robot journalism) to the usage of social bots and the

The Algorithmic Public Opinion: a Policy Overview 4

 creation of deepfakes for propaganda purposes: their role is various,
pervasive and is increasingly institutionalized. In this policy review, I
have identified four significant influential algorithmic systems: content
recommendation, arguably the most influential one as it personalizes
our information environment; automated content analysis tools (content
moderation); social bots; and, finally, the ones employed in news media.

Content Recommendation

With the advent of the Internet, traditional models of information

and news production, distribution and consumption have radically
changed. In order to provide relevant information to users, the epochal
transition from information scarcity to information abundance brought
the need to balance user preferences and algorithmic delegation
(Thurman and Schifferes, 2012). As such, user interests became
more important than content quality or social significance (DeVito,
2017). Such development made personalization very appealing in our
individualistic societies. By using online services, the audience can
exert a greater control over news selection (in theory) and eventually,
can focus on issues that they consider more relevant, which in turn
empowers audiences and erodes the degree of editorial influence
over the public’s issue agenda. This is basically how personalization
algorithms legitimized mainstream social media’s gatekeeping role. The
reality, however, is much more complicated and nuanced.

Everyday people and companies benefit from the personalization

of recommender systems. For example, when Netflix recommends a
movie, Youtube a video, Facebook a stream of information or potential
friends, Spotify music and Amazon related products, a recommendation
algorithm is what performs these tasks. Recommendation and
personalization are entrenched; the functioning of social media sites
is inextricably linked to the quality of their personalization algorithms.
Our time and attention are indeed scarce resources whereas the quantity
of potentially relevant information immense. Personalization systems
therefore perform this fundamental role of knowledge management.
They are both socio-technical systems that mediate and influence
interactions with the economic and social environment as well as
technologies of construction of the self that significantly contribute to
affect in the longer term who we are and who we think we are.

Online news consumption is increasingly relevant in terms of public

opinion’s formation. According to existing research (Newman, 2016),
almost 70 percent of online news users surveyed across 37 different
markets worldwide identified distributed forms of discovery as their
main way of accessing and finding news online, with search and social
media being by far the most influential factors, followed by aggregators,
email and mobile alerts. Their influence is substantial. Consider three
of the most visited websites; on Facebook, the posts encountered by
the average user everyday are circa 350 on about at least 1.500 potential
posts (Backstrom, 2013). Thus, roughly 25% are algorithmically filtered
and prioritized and 75% are ultimately hidden to individual users. On
Youtube, algorithms already drive more than 70% of the time spent in

The Algorithmic Public Opinion: a Policy Overview 5

 the video sharing platform and 90% of the ‘related content’ in the right
bar is indeed personalized. On Netflix, around 80% of viewing hours
come through recommendations and only around 20% from its search
function.

(see → Box 1 The These recommender systems are generally used for personalization.
Reductionism of Profiling Personalization is defined as “a form of user-to-system interactivity
Technologies) that uses a set of technological features to adapt the content, delivery,
and arrangement of a communication to individual users’ explicitly
registered and/or implicitly determined preferences” (Thurman et al.,
2011, p.3). It can in fact be explicit or implicit, that is, it can depend on
user’s requests and/or user’s behavioural data (which is mostly created
unknowingly and unwittingly). Personalization can thus be based on
the individual autonomy of choice or on the algorithmic/platform
delegation to infer one’s personal preferences. Both implicit and explicit
personalization increased dramatically in the last years, though many
websites have acted to make passive forms of personalization the
fastest growing forms (Thurman, 2011). This raises further concerns
about unintended consequences and the reductionism of profiling
technologies.

Personalization algorithms are also employed for microtargeting,

that is predictive market segmentation, or even nanotargeting, namely
delivering ads exclusively to a specific user (González-Cabañas et al.,
2021). These are used for targeted political advertisements, so-called
political microtargeting. These help political parties to identify the
individual voters that it is most likely to convince and, at the same
time, match its message to the specific interests and vulnerabilities of
specific voters. These techniques promise to make propaganda even
more tailored to individual voters, and more effective. So far, they are
primarily used in the United States, but have recently gained popularity
in European countries too (Zuiderveen Borgesius et al., 2018).

Content Moderation

Online moderation systems are methods to sort contributions that

are irrelevant, obscene or illegal in order to ensure that the content
complies with legal and regulatory requirements, site/community
guidelines and user agreements. In recent years, there has been an
expansion in research and investment in automated content analysis
tools (see Shenkman et al., 2021). This has been further accelerated
during the COVID-19 pandemic due to concerns about health risks. They
help to solve the challenge of scale, in particular to inspect or remove
content flagged for hate speech or other objectionable or problematic
content. There are, essentially, two types of automated content analysis:

1. Matching models which are generally well-suited for analyzing

known, existing images, audio, and video, particularly where the
same content tends to be circulated repeatedly and
2. Predictive models that can be well-suited to analyzing content for

The Algorithmic Public Opinion: a Policy Overview 6

 which ample and comprehensive training data is available. Exam-
ples can include whether multimedia contains clear nudity, blood,
or discrete objects.

These algorithms are mostly designed to detect child sexual abuse

material, terrorist propoganda, nudity, hate speech and copyrighted
content. Notably, Facebook utilizes a largescale matching algorithm
on “every image uploaded to Instagram and Facebook” to scan against
an existing curated database of “misinformation,” including COVID-19
misinformation. But there are various other usages; for instance, any
website can use these algorithms to support or automate comment
section moderation. Online social matchmaking services like OkCupid
have utilized algorithms to scan for re-uploads of banned profiles.
Amazon utilizes ‘audio fingerprinting’ to prevent mentions of the word
“Alexa” in advertisements from mistakenly triggering Alexa devices
and resulting in negative customer experiences. In essence, these
algorithmic systems help to detect content with certain features to
eventually remove them.

The overall impact of these predictive systems on the speech

of platform users is still poorly understood. While there are many
important and useful advances being made in the capabilities of
machine learning techniques to analyze content, policymakers,
technology companies, journalists, advocates, and other stakeholders
need to understand and deal with the limitations of these algorithms.
The risk is that they become a form of censorship and eventually they
will lead to other detrimental impacts on human rights, and on the
ability of platforms to function as spaces for discourse, communication,
and interpersonal relation (see Cobbe, 2020).

Social Bots

Bots are automated computer software that perform tasks along a set
of algorithms, and they are at work all over the Internet at varying levels
of sophistication. According to one oft-cited estimate1, over 37% of all
Internet traffic is not human and is instead the work of bots designed
for either good or bad purposes. At the same time, a study showed that
30% of users can be deceived by a bot. A team of researchers from the
University of Southern Carolina and Indiana University released figures
suggesting that between 9% and 15% of active Twitter accounts are bots,
while higher percentages of politician’s followers are also fake. They
can eventually be used for propaganda: russian bots tweeted conspiracy
theories at US president Donald Trump in an effort to get him to spread
the stories through the media.

Social bots legitimate uses vary: crawler bots collect data for search
engine optimization or market analysis; monitoring bots analyze website
and system health; aggregator bots gather information and news from
different sources; and chatbots simulate human conversation to provide
automated customer support. Gorwa and Guilbeault (2020) identify six
main bot types:
1 https://www.imperva.com/resources/re-
source-library/reports/2020-Bad-Bot-Report/.

The Algorithmic Public Opinion: a Policy Overview 7

 1. Web Robots, crawlers and scrapers which download and index
websites in bulk.
2. Chatbots, programs that approximate human speech and interact
with humans directly through some sort of interface.
3. Spambots, bots that can be used to send spam en masse or per-
form DDoS attacks.
4. Sockpuppets and Trolls, bots which can be deployed by govern-
ment employees or regular users trying to influence discussions, to
fabricate reviews and post fake comments about products, people
or institutions.
5. Cyborgs and Hybrid Accounts which are hybrid bots: bot-assisted
human or human-assisted bot.
6. Last but not least Social Bots, bots that automatically produces
content and interacts with humans on social media.

Social bots are likely the most influential one in the context of public
opinion, as they are very common and have a wide variety of uses. The
Department of Homeland Security describes them as programs that “can
be used on social media platforms to do various useful and malicious
tasks while simulating human behavior.” Lutz Finger identified 5 uses
for social bots relevant to the algorithmic public opinion: 1) foster fame:
having an arbitrary number of (unrevealed) bots as (fake) followers
can help simulate real success in a sort of self-fulfilling prophecy. 2)
Spamming: having advertising bots in online chats is similar to email
spam, but a lot more direct mischief: e.g. signing up an opponent with
a lot of fake identities and spam the account or help others discover it
to discreditize the opponent. 3) Limit free speech: important messages
can be pushed out of sight by a deluge of automated bot messages. 4)
Fishing: to fish passwords or other personal data. And finally 5) bias
public opinion: influence trends by countless messages of similar
content with different phrasings (e.g. hashtag flooding or tweetbombing,
to make certain hashtag trending topic) or even create fake grass-roots
movements (also called astroturfing, which is comes from ‘astroturf’
meaning ‘fake grass’, known in the online context as ‘cyberturfing’
(Cobbe and Singh, 2019).

In essence, social bots attempt to ‘game’ the algorithm by inflating

the ‘reputation’ of content and thus increase its likelihood of being
recommended or its position in algorithmic content rankings. By
posting content strategically and artificially inflating views, likes, shares,
and other metrics, networks of bots can together shape the construction
of online spaces. They can therefore artificially seed political messages
in organic discussions, bring greater attention to stories (real or fake),
and boost ideas (fringe or otherwise) into mainstream discussion.

Algorithms in News Media

Several innovative algorithms and ‘AI’ are increasingly used in the

production of media news and journalism for various scopes, especially
in computational journalism (see Diakopoulos, 2014). These generally
help innovate the production and consumption of news and the

The Algorithmic Public Opinion: a Policy Overview 8

 journalistic investigations as well as provide better services at lower
costs. Here I summarize the main algorithms employed in news media
and journalism providing concrete examples.

Automated production of news (robot journalism). The Press

Association in the UK produces more than 30.000 local info per month
using AI tools. In 2015, LeMonde used a technology called Syllabs
during the departmental elections to write local articles on the results
of the 36,000 municipalities and cantons affected by the elections.
The Norwegian News Agency produced sports report 30 seconds after
a football game. AI helps journalists find new angles to a story. For
example, the INJECT project helps journalists scanning articles on a
given topic and makes a proposal for a different angle by suggesting
the story via different actors, submitting related cartoons or data,
challenging by asking incidental questions, or producing facts cards on
the topic.

Investigative journalism. For example, data driven technologies and AI

could provide significant improvements to the in depth analysis of vast
amounts of data of journalistic value such as the Panama Papers or other
leaks. More generally, data mining and processing, and AI can help
journalists to provide more accurate information and develop a more
comprehensive understanding of reality. For example, Serelay is a start-
up which allows verifying if a supposed location of a captured image is
correct and check if no post-processing of this image took place.

Audience analytics. The British news outlet The Guardian has

developed algorithms that allow journalists to understand how
audiences are responding to content - whether the audiences are
reading the texts thoroughly or engaging with it in a particular way.
The Neue Zürcher Zeitung uses a personalized and dynamic paywall
with 150 criteria points determining the ‘hot point’ when a reader can
be converted to a paying version. Similarly, Poool allows for requiring
payment only for certain sections of the online newspaper, to charge
for articles for regular readers or to remove the paywall for occasional
readers. It is based on an in-depth and incremental knowledge of each
reader. The wider application of data and AI tools will enable content
creators and producers to understand the impact of specific content
on the audience and thus make better financing, production and
dissemination decisions.

Automated Translation. AI can also be handy for automated

translation for international broadcasters such as Deutsche Welle or
Arte, allowing them to engage with wider audiences and increase their
reach - although human polishing is still needed for professional quality
translations.

All these innovative techniques and methods are changing (and

promise to change further) news media and journalism. As such, they
fundamentally concur to influence the algorithmic public opinion.

The Algorithmic Public Opinion: a Policy Overview 9

 Concerns Related to the Algorithmic
Public Opinion

There are several concerns related to the rise of algorithms in

the media landscape. This chapter provides a summary of the major
concerns. To begin with, market concentration in digital environments
have led to the concerns that only a couple of algorithms have
unprecedented and unaccountable influence on the information
environment and news consumption of millions if not billions of people
worldwide (algorithmic gatekeeping). Secondly, the black-box nature of
algorithms raise concerns over the challenges of making these systems
transparent, explainable and ultimately accountable (algorithmic
opacity). Strictly intertwined, there is the inevitable risk that algorithmic
systems are biased and lead to a variety of discriminations (algorithmic
bias). Furthermore, there is a long-standing concern that algorithms –
and more generally online social networks – can decrease the diversity
of information to which individuals are exposed to and that eventually
consume (so-called filter bubbles and echo chambers). The highly
personalized data-driven and algorithmically-mediated information
flows also trigger concerns over the risks of amplifying ‘fake news’
(disinformation) and creating addictive patterns of consumption
(algorithmic manipulation). Finally, there are a number of equally
significant concerns that are here briefly mentioned such as favoring a
superficial news consumption, distraction, hate, and narcissism.

Algorithmic Gatekeeping

Nowadays few social media platforms – in particular Facebook,

Youtube and Twitter – dominate most of the informational online media
traffic. It is nowadays agreed in the literature about competition in
social networks that an equilibrium can sustain only a small number
of such intermediaries and a concentrated market structure is thus
expected. Economies of scale in the production of news may indeed lead
to monopolies. This is mainly because of network effects which occur
when the value of a platform to any user increases exponentially with
the number of already present users. As a matter of fact, innovation
in this context has been pretty limited and the history of alternative
social media is a history of failures. This is an inevitable outcome as
long as visibility and scalability depend on economic and political
resources (Fuchs and Marisol, 2015). The result is that few companies
are de facto an oligopoly and, through their algorithms, they developed
and preserve an immense ‘opinion power’, defined as the ability of the
media to influence processes of individual and public opinion formation
(Helberger, 2020). In addition to this, algorithms for content moderation
also allow ‘algorithmic censorship’ (Cobbe, 2020). This power should
come with certain responsibilities and legal liabilities (as it will be
discussed in the section ‘Editorial Obligations and Intermediary Liability’).
How these companies design their platforms, how they allow content to
flow, and how they agree to exchange information with competing
platforms have clearly direct implications for both human rights

The Algorithmic Public Opinion: a Policy Overview 10

 and innovation. The intermediation of content is related to several
conditions of democracy; how people receive news, the articulation
of relationships and associations; access to knowledge, and spaces for
deliberation about issues of public concern. Nevertheless, platforms
have always firmly argued that they are technology companies rather
than media companies, thereby avoiding media regulation and editorial
responsibilities (Napoli and Caplan, 2017). Algorithms, however,
are not their only source of power; social media policies, technical
design choices and business models also serve as a form of ‘privatized
governance’ directly enacting rights and regulating the flow of
information online and, in doing so, promote or constrain civil liberties
(DeNardis and Hackl, 2015)

Distortive effects must be assessed in the context of the economically

oriented algorithms. Especially with regard to social media, the main
imperative behind personalization algorithms is indeed to “increase the
number of regular users and the time that they spend on the platforms—
to increase engagement” (Napoli, 2019, p. 36) because for platforms,
it is primarily important to keep as many users for as long as possible
to—corresponding to their business model—sell ads. There are indeed at
least two main differences between traditional journalistic curation and
algorithmic curation: on the one hand, relevant editorial news values
(such as controversy, negativity, and elite people) interact with each
other. A single factor is never decisive—unlike on Facebook, where only
popularity with the users and their personal network mainly determines
the content of the news feed. On the other hand, the basic direction is
fundamentally different: while the news values that have traditionally
guided journalistic gatekeeping emphasize social significance, the news
values of intermediaries like Facebook focalize personal significance
(see → Box 2 The Case of and are thereby primarily audience-oriented. To fully understand
Facebook’s Algorithm) these dynamics, however, it is necessary a certain level of algorithmic
accountability. This leads to the issue of algorithmic opacity.

Algorithmic Opacity (or ‘The Black-box Problem’)

A primary source of concern is represented by the opacity of

algorithmic systems. The understanding of algorithms in the academic
world is weak due to two major factors: the impermanence of Internet
technologies and the black-boxed nature of most influential algorithms
(see Bodo et al., 2017). The first means that by nature the Internet is
transient, rapidly changing at a rate that usually outpaces the research
process. Algorithms are highly mutable. Google, for example, changed
its algorithm 4887 times in 2020.2 Secondly, the black-boxed nature of
algorithms occurs not only to protect trade secrets but also prevent
malicious hacking and gaming of the system (Pasquale, 2015). Of course,
these features make not only research but also the policy focus on
algorithms very challenging.

Despite their relevance, algorithms remain mysterious, invisible

2 https://www.italian.tech/2021/05/23/news/
le_ricerche_di_google_e_il_senso_della_vita-
and mostly unknown. Not only they are protected by trade secrets but
302295003/?ref=RHTP-BG-I302503236-P6-S3-T1 they are often inscrutable even to their creators. This is the problem

The Algorithmic Public Opinion: a Policy Overview 11

 of interpretability3 (Albanie et al., 2017). Similarly, it is problematic
to unveil the reasons why algorithm made a specific choice. This
is the problem of explainability4.These create structural problems
for algorithmic transparency and accountability. Furthermore,
personalization algorithms adapt to one’s behavior and expressed
preferences and, therefore, they could even be considered unique
to individuals. All in all, algorithms are complex socio-technical
assemblages that are pervasive but still invisible, inscrutable, highly
mutable and often adapted to an individual’s profile or a specific context
or dataset: these features make them an extremely difficult object to
grasp, research and, eventually, to govern. Researchers and civil society
have attempted to intepret the values inscribed into these algorithms
and their consequences, especially through Applications Programming
Interfaces (APIs). This allowed a number of relevant studies. Yet, in the
aftermath of the Cambridge Analytica scandal, social media platform
providers such as Facebook and Twitter have severely restricted access
to platform data via their APIs, what Bruns (2019) ironically called
‘APIcalypse’. This has had a particularly critical effect on the ability
of social media researchers to investigate phenomena such as hate
speech, trolling, and disinformation campaigns, and to eventually hold
the platforms to account for the role that their affordances and policies
might play in facilitating such dysfunction. Alternative data access
frameworks, such as Facebook’s partnership with the controversial
Social Science One initiative, represent an insufficient replacement
for fully functional APIs. Many researchers have argued the need for
a return to web scraping5, and a growing number of practical tools for
scraping data from Facebook, Twitter, and other platforms are now
becoming available. This approach, however, is inherently problematic.
In addition to its dubious legal status, it is ethically questionable. This
institutionalized algorithmic opacity is eventually one of the paramount
concerns: if algorithms cannot be accountable, how can we trust them?

Algorithmic Biases and Discrimination

Algorithmic bias describes systematic and repeatable errors in

a computer system that create unfair outcomes. In the context of
automated content moderation there is an obvious risk of amplifying
biases present in the real world. It is well documented that datasets
3 Interpretability is the degree to which a hu- are susceptible to both intended and unintended biases. How specific
man can understand the cause of a decision or concepts are represented in images, videos, and audio may be prone
consistently predict the model’s result. In other
words, how accurate a machine learning model
to biases on the basis of race, gender, culture, ability, and more.
can associate a cause to an effect. Interpretabil- Automated content moderation perform also poorly when tasked with
ity should not be confused with the concept of
explainability. The former is about being able
decisions requiring an understanding of context and when analyzing
to discern the mechanics without necessarily new, previously unseen types of multimedia. Furthermore, the lack of
knowing why. The latter is being able to quite diversity in engineering and design teams is an additional contributing
literally explain what is happening.
factor to algorithmic bias. All this can eventually lead to discrimination
4 Explainability is the extent to which the
internal mechanics of an algorithmic system can against users. For example, models trained on several of the most widely
be explained in human terms. used hate-speech datasets are up to twice as likely to label tweets by self-
5 Web scraping refers to bots which crawl web identified African Americans as toxic.
pages simulating human Web surfing in order
to collect specified bits of information from
different websites. Content recommendation is also not immune to algorithmic biases.

The Algorithmic Public Opinion: a Policy Overview 12

 Even if the majority of users believes that algorithms are selecting
content neutrally and informing impartially (Gillespie, 2014), the design
of algorithms is inevitably affected by choices made by designers, i.e.,
which factors to include in the algorithm, and how to weigh them.
They indeed use supplied criteria to determine what is “relevant” to
their audiences and worth recommending, though these biases are not
generally recognized.

The most known ones are the following:

Popularity bias

Information intermediaries often include popularity metrics in

their ranking algorithm. A search algorithm, for instance, can give
more weight to information coming from popular websites, to support
majority interests and values. As a result, users may have troubles
finding the less popular and smaller sites, the so-called long tail of
content.

Third-party influence/manipulation

Because the information filtering are automated, they might be

manipulated by activities from third parties. This could occur in several
ways, from clickbait and Search Engine Optimization (SEO) techniques
to ‘bots’ that game social media’s metrics in order to further the spread
of potentially problematic content.

Product/service prioritization

Studies showed that Google and Bing search engines both reference
their own content in its first results position when no other engine does
(Bozdag, 2013). Facebook was also criticized for favoring the products of
its partners. In the last decade, the EU received complaints that claimed
how their traffic drop after Google began promoting its own services
above conventional search results.

Novelty bias

In Google search engine, the number of years a domain name is

registered has an impact on search ranking; domain names that exist for
a period of time are preferred over newly registered ones. In Facebook,
the longer a status update has been out there, the less weight it carries.
A news item is prioritized over an old item. This might, for instance,
lead companies to post updates when their audience is most likely to be
online and using Facebook.

Other biases

The algorithm can also prioritize certain types of information over

others. For instance, the engagement maximization business model
of mainstream social media tends to prioritize ephemeral content to
durable one, short videos (i.e., snippets) to long ones and casual images

The Algorithmic Public Opinion: a Policy Overview 13

 (i.e., snapshots) to written text as they better lock-in users in the so-
called “walled-gardens”.

Media Diversity: Filter Bubbles and Echo Chambers

(see → Box 3 The Challenges of The impact of algorithms on the diversity of information has
Media Diversity) mostly been discussed negatively, assuming that such systems limit
the breadth of viewpoints and topics. The major risk is the creation
of “informational bubbles”: filter bubbles (Pariser, 2011) and echo
chambers (Sunstein, 2017), two sides of the same token. The first is a
kind of cultural and ideological bubble in which an individual continues
to see, listen and read what reinforces its opinions and interests. The
latter is a group situation where established information, ideas, and
beliefs are uncritically spread and amplified, while dissenting views are
ignored. The crucial difference is that the former may not depend on
the user's autonomy and awareness – therefore it is mainly caused by
technological affordances – while the latter pre-exists the digital age and
thus it is primarily driven by social relations.

From an individual perspective, content recommendation might

reduce opportunities to self-determine and could negatively affect
information finding by reducing the exposure to alternative points of
view in the “marketplace of ideas” (Pariser, 2011; Sunstein, 2017) and,
more generally, to serendipitous encounters (Reviglio, 2019). More
generally, the main consequence to provide a ‘too familiar world’ is
that our online life would eventually shift from an intersubjective to a
subjective one (Keymolen, 2016). The consequences may be various:
from the limitation of personal creativity to a reduction in the ability to
build productive social capital (e.g. weak ties).

From a collective perspective, content recommendation can

especially weaken media pluralism. As such, the audience would
become increasingly politically fragmented and polarized and people –
especially the less skilled and literate – more vulnerable to censorship
and propaganda or, better, to “self-censorship” and “self-propaganda”.
This, in turn, would contribute to spread misinformation (Vicario et
al., 2016) and erode interpersonal trust (Keymolen, 2016). It should
be noticed, however, that in nations where political power is divided
among several parties – and not only two like in United States – political
polarization is more difficult to measure, and it is unclear whether it
would even be possible.

There is limited evidence of the existence of these phenomena.

Critics argue that these concerns mostly represent moral panics, and
that personalization could instead foster the cultivation of “expert
citizens” with stronger group identities. Also, they are poorly defined
and, in fact, are used more as generalizing (thus limiting) metaphors
(Bruns, 2019). To this date, research has been often contradictory,
ambiguous and, ultimately, unconvincing (Zuiderveen Borgesius et al.,
2016; Bodo et al., 2017; Tucker et al., 2018).

A number of challenges and concerns, however, persist. Firstly, these

The Algorithmic Public Opinion: a Policy Overview 14
 phenomena are very hard to prove. Most research is often inconclusive
because it is generally survey-based, or is correlational or based on
small or unsatisfactory samples. In light of the fast-changing media
landscape, many studies become rapidly outdated. Nevertheless, the
risks of growing the “digital divide” and “cultural divide” or “epistemic
inequality” remain. Certain privileged group of users, that have higher
(digital) literacy, are able to manage more fruitfully personalization
systems and online information consumption. Instead, a larger group of
users would risk to be exposed only to a minimum, qualitatively inferior,
range of information. Also, the wealthier the social networks, the more
the benefits of personalization, and vice versa.

While insights on the main causes and risks of content

recommendation have been currently understood (see Tucker et
al., 2018), we still lack evidence with regard to the extent of their
consequences. The issue is well beyond simplistic accounts that blame
the algorithms. Information filtering processes take place not only at
the technological level (both algorithms and affordances) but also on
the individual (e.g., selective exposure) and the social (e.g., sharing
practices). And given the vast heterogeneity of users, causes and effects
of content recommendation vary widely.

Algorithmic Manipulation

The manipulation of our perception of the world is taking place on

previously unimaginable scales of time, space and intentionality. There
is indeed plentiful of information wars among several actors — state or
non-state political actors, for-profit actors, media, citizens, individually
or groups (see → Malicous Actors) — to the detriment of citizens
and society at large. The risk of harm includes threats to democratic
processes, including electoral integrity, and to democratic values that
shape public policies in a variety of sectors, such as health, science,
finance and more. The consequent misinformation can lead to public
preferences different to if they were accurately informed, which can
have negative policy implications. The same is true with public opinion
more generally, where policy outputs feed back on public inputs into the
policy‐making process (Dalton and Klingemann, 2007).

Bots, “fake-news”6 and political micro-targeting are the primary

weapons of propaganda and, possibly, manipulation. Social Media
6 Fake-news, intended as “fabricated infor- in particular, and the Internet more generally, face accusations of
mation that mimics news media content in form deteriorating civil debate to the point that facts and truth are now
but not in organizational process or intent” (Laz-
er et al., 2018), is considered by most academics
fertile ground for dispute and subjectivity, while trust with experts and
as a vague buzzword - not to be confused with authorities have decreased, leading to a “post-truth era”. More generally,
overlapping and better defined concepts like: today’s Internet —especially social media (Deibert, 2019)— can be
misinformation – “which is information that
is false, but not created with the intention of manipulative by design threatening individual autonomy (Gal, 2017;
causing harm” – disinformation – “which is Zarsky, 2019). Human behavior can indeed be manipulated by priming
information that is false and deliberately created
to harm a person, social group, organization and conditioning, using rewards and punishments. Eventually, the
or country” – and malinformation – “which is techniques employed can affect individuals’ self-control, self-esteem
information that is based on reality, used to in-
flict harm on a person, organization or country”
and even self-determination. and can stimulate users in a powerfully
(Wardle and Derakhshan, p.20, 2017). subconscious and hormonal way. Facebook’s large-scale emotional

The Algorithmic Public Opinion: a Policy Overview 15

 contagion experiment exemplifies this point (Kramer et al., 2014),
showing how mainstream social media can affect emotions and exploit
vulnerabilities in human psychology.

Since the Cambridge Analytica scandals implicating manipulative and

possibly illegal social media use in Brexit and Trump 2016 campaigning,
challenges and more effective solutions are being discussed. Despite
apparently increasing transparency along with the resultant efforts to
reform, social media often promoted extreme, inaccurate and radical
content—regardless of what malicious actors may do to seed it. For
example, one of the most radicalizing effect is the “rabbit hole effect”
on Youtube, when algorithms capture a user in a spiral of ever more
extreme (more often conspiratorial) content (Yesilada & Lewandowsky,
2021). Manipulation may occur also with political microtargeting,
especially with ‘dark ads’ which are not recognizable as ads at all
(Borgesius et al., 2018). One of the actual risks is that a more or less
certain percentage of ‘persuadables’ – people known to be particularly
vulnerable to targeted messages – can shift elections. In the context
of search engines, this is called Search Engine Manipulation Effect
(SEME) and can shift voters’ preferences by 20% or more (Epstein and
Robertson, 2015).

Then, there is the ability to create compulsion loops which are

found more broadly in a wide range of social media (Deibert, 2019).
These can be triggered through techniques such as variable ratio
reinforcement7 or A/B testing8. Such techniques become more effective
thanks to affective computing (or ‘emotional AI’), captology—the study
of computers as persuasive technologies (Fogg et al., 2002)—and the
emergence of psychographic techniques, along with diverse types of
data such as location-based tracking, real-time data, or keyboard usage.
Consider that suffice only a dozen of Facebook Likes to reveal useful and
highly accurate correlations, such as predicting personality type, even
better than one’s parents prediction (Youyu et al., 2015). In addition,
algorithms can even autonomously explore manipulative strategies that
can be detrimental to users (Albanie et al., 2017).

At the same time, even design choices can be used to implement

deceptive functionalities that are not in the user’s best interest (so-called
dark patterns, see → ‘A Ban to Dark Patterns?’) (Gray et al., 2018). Design
7 Variable ratio reinforcement is a technique
in which rewards are delivered unpredictably.
facets can also intentionally trigger dopamine rushes or other emotional
This unpredictability affects the brain’s dopa- highs, stimulate popularity contests or implicit social obligations
mine pathways in ways that magnify rewards. It (Kidron et al., 2018). Such ability to nudge is defined by Yeung (2018)
occurs when, after X number of actions, a certain
reward is achieved (like in slot machines). In a as a technique of “hyper-nudging” which dynamically configures the
personalized news feed, for example, among user’s informational choice context in ways intentionally designed to
predictably uninteresting content is recom-
mended a predictably ‘serendipitous’ content influence her decisions. As such, hyper-nudging concerns all of the
8 In an A/B testing, the experimenter sets up design process, not only algorithmic decision-making. These kinds
two experiences: ‘A,’ the control, is usually the of nudging techniques are already concerning in the case of negative
current system and considered the “champion,”
and ‘B,’ the treatment, is a modification that at-
effects on children’s wellbeing, including increased risk of suicide and
tempts to improve something—the ‘challenger.’ depression, conflicts with parents and adverse effects on cerebral and
Users are randomly assigned to the experiences,
and key metrics are computed and compared
social development (Kidron et al., 2018).
until it is found the versions that best exploit
individuals’ vulnerabilities.

The Algorithmic Public Opinion: a Policy Overview 16

 There are several other intertwined concerns of the algorithmic
public opinion as it is generally designed nowadays. Above all, the
business model of mainstream social media is driven by the desire to
capture the user’s attention as much as possible. This is often referred
to as the “attention economy”, an economy in which user’s attention is
a commodity to exploit through ever more sophisticated techniques,
including algorithmic ones. These certainly influence individual
behaviors, information consumption and, therefore, the public opinion
in many subtle and often contradictory and counterintuitive ways.

Superficiality and Distraction

Not only it is somehow deterministically assumed that information

overload, multitasking and the disintermediation of information online
are worsening our capacity of problem-solving and critical analysis
(Carr, 2011), but algorithms optimized for maximizing engagement
actually favor endless scrolling of information stimulating superficial
news consumption. One increasingly common strategy employed by the
younger generations to deal with this new environment, for example,
is “news grazing”, which occurs when users do not purposefully
devote time to consuming news, maintaining a passive eye toward
information, or keeping news sessions shorter. It is a form of news
consumption motivated by the need for a wide breadth of information,
rather than depth. It could be an adaptation to information overload.
Such engagement might well be superficial if news consumers stop at
just reading news headlines and short summaries. Algorithms can be
designed to exploit this vulnerability and prioritize ephemeral content to
durable one, as these are more engaging overall.

Narcissism and Online Participation

The quantification and maximization of social interactions contribute

to create a “culture of performance” (Castro, 2016) that seems to be
negatively correlated with well-being (Verduyn et al., 2017). Platforms
like Facebook and Instagram do not manifest much as a public space
but falls within the frame of the private and the exposure of the self.
Among various consequences, narcissism thrives could lead to a de-
politicization of society (Byung-Chul Han, 2016). Either way, it could
lead to showing off political activism. A paradigmatic example is
“slacktivism”, which refers to the act of showing support for a political
or social cause with very limited involvement required nor concrete
effort. The main purpose is boosting the egos of participants in the
movement. Popularity is the omnipresent tacit dream. In this rewarding-
system machine, many - if not most - of the users show some compulsive
behaviors. For example, many experience the fear of missing out from
updates (what is called FOMO) so they would feel anxious without
scrolling, login or internet connection. And despite such compelling
engagement, the resulting environment is eventually toxic and exclusive,
one that even suppresses the participation of most users. The majority
of social media users are indeed lurkers, mere watchers of the online
world, often falling into a “spiral of silence” in which they do not
publicly express their opinions (Noel-Neumann, 1984). Platforms tend to

The Algorithmic Public Opinion: a Policy Overview 17

 favor who plays their rules, and nudge users to play them.

Hate and Divisions

Incendiary and polarized content as well as hate speech are arguably

widespread in social media. The consequences of this environment can
be complex and various; from the dissemination of hate and incivilty
to cultivating a subsectibility to indignation, from stimulating “on-
demand politics” (where politicians deal with constant coverage and
people’s demand) to “identity politics” (when political agendas are
based upon specific identities), from creating “culture wars” (cultural
conflicts between social groups and the struggle for dominance of their
values, beliefs, and practices) to ultimately leaving people distracted,
divided and demoralised. If not “fixed”, social media in the next years
may increasingly facilitate identity-based violence. The risk, in fact, is
that identity-based beliefs tend to eclipse truth-seeking because of the
overriding need to belong and feel morally superior.

Main Actors in the Algorithmic

Public Opinion

Before developing policy solutions focused on the prevention of

negative consequences of algorithms it is paramount to make these
systems transparent. But to make transparency unfold in algorithmic
systems it should take into account not just code and data but an
assemblage of human and non-human actors (Ananny and Crawford,
2018). This means to account for all the relevant actors in the
algorithmic public opinion, most importantly users, policy-makers,
governments, mainstream platforms and non-mainstream platforms,
programmers and designers, moderators, influencers, journalists and
news media, academics, malicious actors etc.. Of course, to maintain
this policy review short, I present here only the main actors: policy-
makers, platforms, users, news media, academics and malicious actors.

Policy-makers

To regulate or not to regulate? This is likely the main dilemma

of policy-makers. The rapid development of the media landscape
has notably left regulators to have to consistently catch-up with new
innovations and inevitable trade-offs and conflicting human rights.
This reality often leads to the “collingridge dilemma”; this argues that
it is relatively easy to intervene and change the characteristics of a
technology early in its life cycle. At this point, however, it is difficult
to predict its consequences and regulators need convincing evidences
before acting. Later, when the consequences become more visible and
evidences are collected, it may be much more difficult to intervene. This
is particularly true if we are dealing with the novelty and unprecedented
scale of most algorithms in this context.

The Algorithmic Public Opinion: a Policy Overview 18

 When a consensus to intervene is found, policy-makers are also
faced with two interrelated questions: who should regulate and how
these should be regulated. In general, there are three main approaches:
traditional regulation, self-regulation and co-regulation who set and
enforce different regulatory goals, standards and justifications (Hirsch,
2010: Finck, 2018). These reveal different approaches to regulation, yet
they operate on a spectrum. Also, all of them are already set in place
to some extent. On the one hand, platforms are already self-regulating
entities; they determine the terms and conditions of their intermediary
function and define online and offline standards of behaviour. On the
other hand, there is an information asymmetry between platforms and
policy-makers that naturally requires some forms of co-regulation which
usually takes the form of voluntary codes of conduct and negotiated self-
regulatory agreements. The state role, in this case, is relatively limited to
more of an informal oversight and steering role.

Fundamentally, policy-makers lack a shared vocabulary or

frameworks for approaching the incoming challenges, and new
models for digital governance are likely needed to be developed.
The algorithmic public opinion is in fact particularly challenging
to policy-makers as it intersects with a transformation of media
itself, due to media convergence, media globalization and the rise
of global digital platforms. Traditional policies for broadcasting,
telecommunications, and media are more often inadequate. There is
a risk that “regulatory transference” – the application of existing law
and regulation to new business models and market conditions – does
not address the right problems, and has unintended consequences.
Legal and regulatory frameworks designed for one set of market and
technological circumstances may be ineffective or inappropriate in
others. The complexities of contemporary digital systems and networks,
algorithmic content filtering, data ecosystems, social media, and
cross-platform activities necessitate different methods to address the
issues and challenges they pose. Domestic policies can address some
issues, but global policy is progressively more significant to address
communication challenges and classic regulatory concerns such as
compliance, enforcement, and efficacy.

In the past years policy-makers have undertaken several different

initiatives to regulate social media platforms and their algorithms.
In particular, this has included horizontal instruments, such as
competition and data protection law, which are not specifically tailored
to social media and their algorithms, but may still have some spillover
benefits for certain purposes. Antitrust and consumer protection law,
in fact, already place certain limits on platforms, for example limiting
their ability to prioritize their own services without discriminating
other actors or even forbidding covert advertising, which means that
platforms have a duty to disclose whether content is being sponsored.

To date, (supra)national governments have tried some forms of co-

regulation, such as the EU Code of Practice on Disinformation, in which
the major tech companies – Facebook, Google and Twitter – pledged
to work more actively to lessen the spread of disinformation and hate

The Algorithmic Public Opinion: a Policy Overview 19

 speech online. These, however, are usually non-binding and the rules
could be interpreted rather loosely by companies. Over the years, many
critics in academia and civil society have argued that co-regulation and
self-regulation attempts do not provide sufficient incentives to act in the
public interest. It is nowadays widespread these need to be replaced by
the law (Floridi, 2021b).

Policy-making also presents a number of structural limitations that

limits its decisive role. To begin, there are legitimate but concerning
lobbying activities to take into account. Even the often praised European
General Data Protection Regulation (GDPR) is one of the most lobbied
pieces of EU legislation to date (Edwards and Veale, 2017). Lobbying is
very strong even for new European legislative drafts such as the Digital
Services Act and the Digital Markets Act. The resources these companies
can employ to negatively affect innovative policies are huge. This brings
concrete risks that need further analysis to, eventually, cultivate a more
effective public oversight. Similarly, another phenomenon is the one
of revolving doors which refers to the movement of personnel between
roles as legislators and regulators on one hand, and members of the
industries affected by the legislation and regulation on the other. For
example, The Google Transparency Project has so far identified 258
instances of revolving door activity (involving 251 individuals) between
Google or related firms, and the federal government of US, national
political campaigns and Congress during the Obama’s presidency.
Bank et al. (2021) argue that these should be simply blocked. Finally,
there are political pressures that might worsen deliberation. Notably,
what Ananny and Gillespie (2016) call “governance by public shocks”.
They argue that social media regulation is often driven by “public
shocks”, these are shocks that “sometimes give rise to a cycle of public
indignation and regulatory pushback that produces critical - but often
unsatisfying and insufficient - exceptions made by the platform (p.3).”
Such reactive approach from politics works against proactive, long-term
policy solutions.

Mainstream Platforms

Today’s social media landscape is characterized by an oligopolistic

market in which a small group of platforms — above all Alphabet
(Google) and Meta (Facebook)— act as the ultimate gatekeepers for
billions of users worldwide. The European Digital Service Act refer
to these as “very large online platforms”, the ones with more than 45
million monthly active users, providing special rules. When we talk
about platforms we in fact refer to the these companies which provide
other than Google search and Facebook other fundamental online
services such as Google maps, Google news, Youtube, Whatsapp,
Instagram and many others.

These companies have created a global architecture of data capture

and analysis produces rewards and punishments aimed at modifying
and commoditizing behaviour for profit. Basically, they harvest
large amounts of data (i.e. big data) to identify patterns of consumer

The Algorithmic Public Opinion: a Policy Overview 20

 behaviours, preferences, tendencies, etc. This business allows the
existence of “free” services. In order to access any social service
platform, however, users have to give up the social data they generate
and accept to be surveilled to help algoritms predict their desires
through personalized ads. It is indeed the ad-industry the main source of
revenue of these companies. In 2017, advertising constituted 87 percent
of Google’s total revenue and 98 percent of Facebook’s total revenue.
This is what is often referred as ‘surveillance capitalism’ (Zuboff, 2015)
in which the capitalistic logic of accumulation produces “hyperscale
assemblages of objective and subjective data about individuals and
their habitats for the purposes of knowing, controlling, and modifying
behavior to produce new varieties of commodification, monetization,
and control” (Ibid., p.85). There is indeed an economic and political
antagonism between users’ interest in data protection and corporate tax
accountability on the one side, and corporations’ interest in user data’s
transparency/commodification and corporate secrecy on the other side.

Mainstream platforms provide a personalized experience, but

very little control over information filtering processes. This control
asymmetry perpetuates an epistemic imbalance in which platforms
know people more than people know themselves. Many of the practices
involved in surveillance capitalism are challenging social norms
associated with privacy and, thus, are contested as violations of rights
and laws. Consequentially, these corporations have learned to “obscure
their operations, choosing to invade undefended individual and social
territory until opposition is encountered, at which point they can use
their substantial resources to defend at low cost what had already
been taken” (Ibid., p.85). On their side, mainstream platforms defend
themselves through a mix of trade secrets, economic claims, promises
of self-regulation, and technological solutionism, forestalling real public
oversight. They essentially respond to regulatory threats by claiming
that:

1. Their systems use proprietary knowledge that they cannot

publicly disclose;
2. Their business models require large-scale data harvesting;
3. People are unwilling to pay for services that are currently ensured
by people’s data;
4. Encryption technologies, transparency commitments, and con-
trolled data disclosure obviate the need for public oversight.

It is important to understand why and how such model became

dominant. According to Zuboff (2015) there are a variety of reasons.
Firstly, this system was constructed at high velocity and designed to
be undetectable. Structural asymmetries of knowledge and rights, in
fact, made it impossible for people to learn about the above practices.
There was no historic precedent, so there were few defensive barriers
for protection. Leading tech companies have been over-estimated,
in a way respected and treated as ‘emissaries of the future’. On the
other hand, individuals quickly have become to depend upon the
new information and communication tools as necessary resources,

The Algorithmic Public Opinion: a Policy Overview 21

 requirements – at times even preconditions – for social participation.
As Zuboff argues: “the rapid build-up of institutionalized facts (…)
produced an overwhelming sense of inevitability (p.85).” And as the
philosopher Mireille Hildebrandt (2015) rightly claims: ‘the temptation
to accept that things are the way they are because technology is the way
it is, has a strong hold on public imagination […] once a technology has
consolidated it acquires a tenacity that is not easily disrupted.’ (p. 174).

This model should lead – and to some extent is indeed leading

- policy-makers to consider more radical policy solutions to tame
the unprecedented, institutionalized, platforms power. Considering
that these companies represent key nodes in economic, social, and
informational flows it has been argued that they could even be treated
as public utilities (Rahaman, 2018) while the data generated by users as
a public resource (Napoli, 2019). Of course, it is not that simple; there
are a number of fundamental and complex systemic issues related to the
power of these companies that will be introduced and further discussed
in the next Chapter.

Users

Users are the ultimate gatekeeper. This is why it is mostly stressed

that for any concern related to the algorithmic public opinion what is
most needed is literacy; not only media and algorithmic literacy but
also digital skills, including an awareness of the challenges of digital
societies. Users can also become active players, helping platforms and
authorities in many ways: from flagging content to self-organized troops
against disinformation online and, more generally, demanding more
transparency and autonomy of choice.

The algorithmic system which is arguably the one in which users can
act more substantially upon is certainly content recommendation. In
this context, it has often been stressed the primary role and therefore
responsibility of users as a prominent argument to argue against strong
regulation. Many of the problems with media pluralism are indeed
mostly user-driven (Helberger et al., 2018). It is therefore necessary
to understand users, their behavior and their perceptions. These,
in fact, deeply influence their trust on platforms and, in turn, how
they use them. Generally, users across the world seem to embrace
personalization algorithms as they enjoy the services they offer on a
daily basis. (Newman et al., 2016). As said, for the majority of users
news algorithms are not doing anything wrong: they select neutrally
and inform or recommend impartially (Gillespie, 2014). As such, most
people trust more themselves in delegating to algorithms than they trust
journalists. Users indeed guess the functioning of algorithms through
what are called “folk theories”, developing a more or less accurate
“algorithmic imagination” (Bucher, 2017). However, most users still
miss the fact that there is no objectivity in the realm of algorithms.
Personalization algorithms embodies specific forms of power and
authority. Still, most users have no reservations in principle against
news being distributed through AI-driven tools (Thurman et al., 2019).

The Algorithmic Public Opinion: a Policy Overview 22

 Yet, they increasingly tend to be worried that “more personalized news
may mean that they miss out on important information and challenging
viewpoints” (Newman et al., 2016, p.113). Research indicates that they
expect to play a more active role in the interaction with AI-driven tools
such as personalization systems (Helberger et al., 2019). Also, there are
concerns about being wrongly profiled, as well as privacy concerns.

In short, the majority of users trust algorithms but less and less and
they enjoy every day their fundamental services more and more. Though
they can have some reservations about algorithms, many have even
more reservations about journalism and editorial selection. Indeed,
journalism is in crisis and more often survives with click-baiting. As
Internet critic and scholar Evgeny Morozov (2017) puts it “what it has
gained in profitability, it seems to have lost in credibility.” Trust is
fundamentally at stake in these processes. Nevertheless, perceptions,
expectations and behaviors over (news) content recommendation
remain rather nuanced, ambivalent and at times even inconsistent.

There are two main explanations for the widespread default trust and
lack of individual and institutional responsiveness over platforms and,
in particular, their opaque and limitedly controllable personalization
systems. On the one hand, as said above, one compelling explanation is
that nothing similar to ICT development ever happened in the past, so
there were few institutional and individual defensive barriers (Zuboff,
2015). Societies quickly came to depend on these new information
and communication tools as necessary resources, and at times
even as preconditions for social participation. On the other hand,
users’ behavior has to be analyzed in light of behavioral economics
and psychology (Acquisti et al., 2015). Drawing from the work of
Kahneman (2011), Thaler and Suntein (2019) describe how our bounded
rationality affects how we assess the likelihood of future events and
how our individual biases and vulnerabilities could be exploited. This
particularly explains the so-called “privacy paradox” – where people
claim to value privacy but they don’t protect it proactively – yet it can
be applied in many other circumstances such as individual control over
personalization systems or vulnerability to dark patterns.

More generally, the use of AI-driven tools fundamentally alters the

agency of users regarding the news individuals consume. AI-driven tools
have introduced observable measurements of user interaction allowing
detailed insights into audience preferences, impossible to obtain in
non-digital media. Even if users are at the center of these processes,
this increase in user agency is unidimensional as it is solely focused
on observable engagement like clicks or time-spent. And as explained
in the previous chapter, users do not know what information they are
automatically excluded from. The challenges outlined above lead to a
fundamental open questions: can users be sufficiently informed to be
fully responsible in such a fast-changing complex media environment?
To what extent and in which forms they ought to be supported?

The Algorithmic Public Opinion: a Policy Overview 23

 News Media

News Media and its adaptation to the digital era is fundamental to

understand the actual algorithmic public opinion. It is indeed well-
known that news media and journalism are in a serious economic and
trust’s crisis. Over the last two decades, as people increasingly began
to consume news through smartphones and other electronic devices, a
shift in the consumption of print media channels has occurred. News
from a variety of online sources, like blogs and other social media,
resulted in a wider choice of official and unofficial sources (e.g. citizen
journalism), rather than only traditional media organizations. One of
the consequence of this is that most people expect online news to be
free and do not buy newspaper anymore. News media – and especially
newspapers – have therefore seen print revenues heavily decreasing,
eventually reducing their staff and coverage. News organizations have
been challenged to recover losses and find new business models for
making journalism sustainable.

The industry is struggling and dealing with this crisis in different

ways. One initial widespread reaction has been favoring quantity over
quality and sensational titles (so-called click-bait). Through these
practices, news media earned the online information industry the
epithet of “Clickbait Media” (Munger, 2020). Another approach has
been the one of paywall, a method of restricting access to content,
with a purchase or a paid subscription. Beginning in the mid-2010s,
newspapers in fact started implementing paywalls on their websites as a
way to increase revenue, partly due to the use of ad-blockers. While this
approach has its merits, especially in recovering lost revenues, critics
argue that erecting paywall restricts equal access to the online public
sphere.

One of the consequences of the current news media crisis has

been a decline in trust. In recent decades, trust in news has indeed
declined in many parts of the world (Newman, 2020). This decline in
public confidence in the press is part of a broader skepticism that has
developed about the trustworthiness of institutions more generally
– state, science and interpersonal – leading to an overall concerning
trust recession. This has led to discussions on how journalists define
and enact their democratic role and how news organizations ought to
give journalists the freedom and encouragement to engage in trust-
building experiments to facilitate discussion, building community, and
partnering with the public.

As prerviously argued, news media is also confronted with the

innovations and challenges brought by algorithms. Notably, these
have led to data-driven journalism (or data journalism), a journalistic
process based on analyzing and filtering large data sets for the purpose
of creating a news story. Data journalism reflects the increased role that
numerical data is used in the production and distribution of information
in the digital era as well as the increased interaction between journalists
and several other fields such as design, computer science and statistics.
Deciding what’s news is indeed increasingly influenced by quantitative

The Algorithmic Public Opinion: a Policy Overview 24

 audience measurement techniques. In particular, the constant audience
measurement returns real-time and personalized data on practices
consumption of the public. Data are therefore often seen as a primary
avenue towards the sustainability of journalism in the digital era.

The result of this comprehensive datafication process applied to

journalism manifests itself in at least four forms (cfr. D6):

1. Journalistic outcomes are increasingly a byproduct of algorithms

in terms of positioning in the search engines and in the timelines
of social platforms. Widespread practices of search engine optimi-
zation (SEO) makes journalism adapt to the algorithms’ will.
2. The metrics of engagement with journalistic content are available
to everyone (including readers): how much an article is shared,
how many likes it receives, what is its circulation. This can have
different, mixed implications: from deepening an issue that inter-
est the audience to adapt their writing style based on users’ com-
ments. In any case, it implies a more audience-driven journalism.
3. Journalistic metrics are a product of a larger analytics processing
system that monitors individual and aggregated behavior (i.e. Goo-
gle Analytics, Charbeat, Newsbeat, Parse.ly). As Diakopoulos (2016)
argues, this has led to an increase in directly related metrics to the
growth of digital platforms in which the act of consuming news
generates a different transmission of data.
4. Journalistic metrics as a product of behavioral processing of net-
work users (such as NewsWhip, Crowdtangle, Ezyinsights). These
softwares monitor what is happening on the Internet, keep track
of social media signals, monitoring tweets, shares and comments.
These new forms of journalism are not fully understood yet, but
they significantly concur to the development of an algorithmic
public opinion as here conceived.

Academia

Academia has a special role in the algorithmic public opinion.

Being at the forefront of the understanding of the digital revolution,
it helps to raise civic awareness, providing theoretical critiques,
scientific evidences, engaging in public advocacy and do consulting to
policy-makers and businesses. Its independence and effectiveness are,
however, questionable.

Scholars have developed and have been involved in plenty of

meaningful initiatives; building new tools. For example, Twitter Capture
and Analysis Toolkit (TCAT), developed at the Digital Methods Initiative
(DMI) at the University of Amsterdam; providing public oversight on
various issues, from social bots to misinformation. Another example
is the Computational Propaganda Project by the Oxford University which
provides meaningful updated analysis and data for anyone interested;
providing critical insights on new legislative proposals. Another
paradigmatic initiative is the Digital Services Act (DSA) Observatory of the

The Algorithmic Public Opinion: a Policy Overview 25

 Institute for Information Law of the University of Amsterdam, a project
exclusively dedicated to the monitoring of the proposed new European
law DSA; finally, scholars can be part of research and ethics boards,
notably the philosopher at the Oxford University Luciano Floridi was
initially part of the Google AI Ethics Board that was subsequently shut
down.

Building on increasingly unprecedented large datasets, gathered

through the Application Programming Interfaces (APIs) of mainstream
platforms, social media research had proven a particularly fertile
field: the scholarly community developed and shared increasingly
sophisticated methods and tools for gathering, analysing, and visualising
social media data. There is a vast potential for social sciences. Yet, as
previously mentioned, in the aftermath of the Cambridge Analytica
scandal, social media platform providers such as Facebook and Twitter
have severely restricted access to platform data via their APIs, what
Bruns (2019) ironically called “APIcalypse”. This has had a particularly
critical effect on the ability of social media researchers to investigate
phenomena such as hate speech, trolling, and disinformation
campaigns, and to eventually hold the platforms to account for the
role that their affordances and policies might play in facilitating such
dysfunction.

The question of API’s and data access is paramount for researchers

and it’s becoming a matter of increasing contention. Over the past
year, dominant platforms such as Facebook have repeatedly interfered
with independent research projects, prompting calls for reform. In
October 2020, before the US elections, Facebook tried to shut down
an independent audit of their political advertising by NYU. Similarly,
it has retaliated against data collection by the NGO AlgorithmWatch,
sending them threats of legal action on the grounds that independent
data collection violated the platform’s Terms of Service. Given the
considerably asymmetrical distribution of resources and control over
platform affordances between providers and scholars, it is therefore
difficult to see this as a race that can possibly be won by researchers,
even if the legal, moral, and ethical justifications for entering into it can
be found.

Scholars too, similarly to policy-makers, can be tempted by

mainstream platforms in “revolving doors” or somehow hushing them.
Over the years, the platforms have also hired a substantial number of
university graduates – in instrumental fields such as computer science
but also from more critical disciplines such as media, communication,
and cultural studies – to bolster their workforce in key areas of
operation; at times, they also collaborate with external research
teams. And from time to time, platforms offer ‘data grants’ and similar
competitive schemes that explicitly invited scholars to apply for funding
and data access for particular research initiatives. The role of academia
for the understanding, development and governance of the algorithmic
public opinion is undoubtedly significant and multifaceted. Yet, the
potential to collect meaningful evidences is seriously limited by data
access. Furthermore, the essential role of informing stakeholders makes

The Algorithmic Public Opinion: a Policy Overview 26

 scholars critical actors in this context.

Malicious Actors

The algorithmic public opinion has favored new opportunities

for malicious actors; many kinds of profit-driven and especially
politically interested actors are finding ways to exploit social media
and algorithms, in particular to spread propaganda, a phenomenon
often reffered to as ‘the weaponization of social media’. Operations can
be conducted by government agencies, politicians and parties, private
companies, influencers and citizens. Foreign influence operations,
disinformation, and state-sponsored trolling and harassment have
indeed already undermined human rights and degraded the quality of
political news in circulation.

Since 2016 the Computational Propaganda Project of the Oxford

University has monitored the activity of “cyber troops”, defined as
government or political party actors tasked with manipulating public
opinion online (Bradshaw & Howard, 2017). They have examined
the formal organization of cyber troops around the world, and how
these actors use computational propaganda for political purposes. As
such, they have built an inventory of the evolving strategies, tools, and
techniques of computational propaganda. The phenomenon is indeed
concerning and steadily growing. Yet, many of these operations are
almost certainly not even been publicly documented.

In their 2020 report, Bradshaw et al. have classified the valence

and messaging strategies used by cyber troops into four categories:
pro-government or pro-party propaganda, attacking the opposition or
mounting smear campaigns, suppressing participation through trolling
or harassment, drive division and polarize citizens. These strategies are
conducted throughout a number of techniques:

1. The creation of disinformation or manipulated media. This is the

most prominent type of communication strategy and includes cre-
ative so-called “fake news” websites, memes, images or videos, and
other forms of deceptive content online even deepfakes.
2. Dark ads. These are data-driven strategies to profile and target spe-
cific segments of the population with political microtargeting.
3. Trolling, doxing or online harassment. In several countries have
been found evidence of trolls being used to attack political oppo-
nents, activists, or journalists on social media.
4. Censor speech and expression through the mass-reporting of
content or accounts. Posts by activists, political dissidents or jour-
nalists can be reported by a coordinated network of cyber troop
accounts in order to game the automated systems social
media companies use to flag, demote, or take down
inappropriate content.

Cyber troops normally use accounts to spread computational

The Algorithmic Public Opinion: a Policy Overview 27

 propaganda, for example: automated accounts (or political bots),
which are often used to amplify certain narratives while drowning
out other; human-curated accounts, which might use low levels of
automation but also engage in conversations by posting comments or
tweets, or by private messaging individuals via social media platforms;
hacked, stolen, or impersonation accounts (including groups, pages,
or channels), which are co-opted to spread computational propaganda,
though they represent a small portion of the total. More generally, these
cyber troops could also employ so-called click-farms in which are low-
paid workers hired to click on ads but also to generate likes, followers
and, more generally, engagement.

The algorithmic public opinion also entails cyber-espionage and

‘information warfare’ . In this context, for example, bots can also be
used to extract data through web scraping which simulates human Web
surfing in order to collect specified bits of information from different
websites. These techniques can be used for “open-source intelligence”
(or OSINT). OSINT is produced from publicly available information that
is collected, exploited, and eventually disseminated to an appropriate
audience for the purpose of addressing a specific intelligence
requirement. This could, for example, result in sensitive lists that can be
used as a form of political intelligence, as it was recently the case with
a Chinese company who created a list of political important networks
in Western countries simply crawling publicly available information
(Balding, 2020). Furthermore, malicious actors can buy malwares in
the black market (or even totally legit ones) to monitor people, notably
activists, journalists, and political leaders. The recent Pegasus Project
leak is a paradigmatic example (Amnesty International, 2021). Several
governments have been accused to have exploited these malwares.
These malicious tools further threaten freedom of speech. Anyone could
monitor and threaten people who could or would speak up about issue
of public interest.

Other serious concerns are data breaches and how these could be
exploited by data brokers - companies that collect consumers’ personal
information and resell or share that information with others. In the
hands of malicious actors, breached data can indeed become a tool for
cyber espionage, political campaigns and, more generally, information
warfare such as disinformation operations, especially during elections.
This raises serious concerns on the effectiveness – or even enforceability
– of data protection regulation at a global level. Data brokers in fact can
proactively obfuscate the source of their data, making it difficult for any
individual to retrace the paths through which their data were collected
(Reviglio, 2022).

The Algorithmic Public Opinion: a Policy Overview 28

 The Governance of the Algorithmic
Public Opinion

The governance of the algorithmic public opinions entails the

challenges and opportunities to re-design our societies. In fact, to re-
think how information is produced and disseminated, how we mediate
our relations online, how we get news and form opinions individually
and collectively and, ultimately, how to increase the quality of public
opinion and, therefore, of democratic deliberation. These represent
fundamental issues for preserving democracies and human rights for
the years to come. Indeed, the fast-paced media environment along
with disruptive technological innovations (prominently artificial
intelligence) and a number of epochal challenges (from climate change
to increasing geopolitical tensions) require regulators, companies and
civil societies to be able to continuously maintain cohesion and reach
new agreements. This also requires finding agreements on new set of
rules to develop a more innovative, transparent and inclusive Internet
and social media environment. The following analysis, therefore,
aims to help to shed light on the limitations and challenges of current
regulations and policies. Firstly, the most critical debates for the
governance the algorithmic public opinion are introduced. Secondly,
relevant policy challenges and debates are mapped; on the one hand,
systemic issues and the most prominent policy approaches for the
governance of algorithms, antitrust, data protection and media liability
are briefly introduced. On the other hand, more specific issues such
as recommender systems, content moderation, social bots, political
microtargeting and disinformation are discussed.

The Challenges of Governing the Algorithmic

Public Opinion

For emerging digital technologies, legislation is a rather blunt

tool; it always risks being too broad to be useful, or too specific to be
future-proof. The Internet has also undermined the benchmarks that
acentury of international cooperation helped to build; the principles
of territoriality, universality of values and effectiveness in the
cyberspacecollide with the fluidity of data, algorithms and information
flows. Social media and their personalization algorithms are certainly
not easy to regulate either; the question of regulating online information
flows is part of a complex regulatory system involving many different
legislative acts and normative approaches.

The question of how to effectively govern platforms and their

algorithms is an open challenge. Together with the centrality of
national authorities, it is widely shared the need to move away from
homogenous top-down models towards decentralized, reflexive,
effective, collaborative and cooperative frameworks that are ‘polycentric’
(see Finck, 2018). Regulatory conversations on social media platforms
are indeed already polycentric in that they are transnational and
multisectoral. Polycentricity is indeed inherent to new governance

The Algorithmic Public Opinion: a Policy Overview 29

 models where no actor does make decisions unilaterally. Unlike
traditional conceptions of law that rely on a unitary source of authority,
new governance is based upon a dispersal and fragmentation of
authority, and rests upon fluid systems of power sharing. In this context,
it is essential, for example, to be able to combine digital regulation with
ethics and governance. The regulation of the digital, the governance
of the digital and the ethics of the digital (whether computer, AI,
information or data ethics) are in fact different normative approaches,
complementary, but not to be confused with each other (Floridi, 2018).
Indeed, not every aspect of regulation is a matter of governance and
not every aspect of governance is a matter of regulation. For example,
governance may comprise guidelines and recommendations that
overlap with, but are not identical to regulation, whereas ethics shapes
regulation and governance through the relation of moral evaluation.

Platform Governance

Aside from theoretical approaches, there is also a more pragmatic

need to coordinate regulatory efforts while innovating the governance
of platforms. Importantly, “platform governance” - understood as
the set of legal, political, and economic relationships structuring
interactions between users, technology companies, governments, and
other key stakeholders in the platform ecosystem (Gorwa, 2019) — is
indeed rapidly moving away from an industry self-regulatory model
and towards increased government intervention (Helberger, Pierson, &
Poell, 2018; Floridi, 2021b).

As a matter of fact, platform governance does not depend exclusively

on a single source of accountability or regulation, but rather on more
complex and multistakeholder systems of governance. The complexity
of governing the algorithmic public opinion, in fact, results also from
the fact that a large number of different stakeholders are involved in
the development, production, distribution, exploitation and marketing
of social media content. The three main categories at stake are users,
content providers/producers and distributors/platforms, which in
turn are split into a number of different types of actors – much more
diverse in the digital environment than in the analogue environment.
The regulatory space has indeed increased dramatically due to the
borderless nature of the digital world, as it has the technical expertise
needed to create effective, appropriate and enforceable rules. For
these reasons, the involvement of various stakeholders in regulatory
approaches has become ever more important.

Multistakeholderism

One way to mitigate the challenges of the governance of the

algorithmic public opinion is, therefore, to incorporate multistakeholder
and co-regulatory elements. On the one hand, both governments and
international organisations have often failed to produce adequate
solutions to contemporary corporate transnational governance and
policy issues. Traditional regulation - where governments seek to make
corporations comply under threat of legal and financial penalties - it

The Algorithmic Public Opinion: a Policy Overview 30

 is not always easy to pass, as industry lobbies heavily to protect its
interests, and even once rules are in effect, ensuring compliance —
especially when firms are headquartered in different jurisdictions
— is no easy task. Thus, multistakeholderism becomes meaningful;
different governance stakeholders have differing levels of regulatory
capacity that they bring to the table: each type of actor has different
competencies that are required at different phases of the regulatory
process, from the initial agenda-setting and negotiations to the
eventual implementation,monitoring, and enforcement of governance
arrangements.9

“Cooperative Responsibility”

Helberger, Pierson and Poell (2018) argue that “the realization of

public values in platform-based public activities cannot be adequately
achieved by allocating responsibility to one central actor (as is currently
common practice)” and therefore envisage a “dynamic interaction
between platforms, users, and public institutions”. Social media policy
is in fact moving towards a ‘cooperative responsibility’, that is, the
result of the dynamic interaction between platforms, users and public
institutions. In such (re)distribution of responsibilities, they identify
four key steps:

1. To collectively define the essential public values at play in particu-

lar contexts
2. Each stakeholder (platforms, governments, users, advertisers, and
others) accepts that they have a role to play in the realization of
these value
3. It is developed a multi-stakeholder process of public deliberation
and exchange, in which agreement can be reached.
4. They translate the outcome of public deliberation and agreements
into regulations, codes of conduct, terms of use and, last but not-
least, technologies (e.g. ‘by design’).

Towards a New Governance

Co-regulatory, multistakeholder solutions that allow the dynamic

constructive relationship between ethics, regulation and governance
bear the potential to allow for more informed decision-making, easier
enforcement, and continuous and accountable review and assessment.
The experimental nature of this process allows for mutual learning and
the identification of best practices as well as for a dynamic adaptation of
the relevant rules over time. However, to believe that “a single effective
9 It is also true that the governance of online
content on platforms is a far less multistake- and proportionate regulatory approach could be designed in such a
holder than the typical Internet Governance way as to tackle every one of these matters is highly presumptuous and
(IG) of internet protocols and standards, with
far fewer formalised institutions and fora. To neglects the wide array of complex social factors underpinning the
be clear, multistakeholderism is no panacea production, sharing and engagement of such content (Nash, 2019, p.19).”
and civil society has been often marginalised in
IG, eventually merely serving to legitimise the
This review, in fact, has no ambition to provide any comprehensive
process for other, more powerful actors. governance approach but only to stimulate the policy debate by

The Algorithmic Public Opinion: a Policy Overview 31

 discussing promising normative approaches and policies proposals.

Systemic Issues

Digital media and the growing data economy has profoundly

disrupted media markets, challenged the existing business, production
and consumption of news, and presented fundamental challenges
for pol cy-makers. The increasingly intertwined areas of algorithmic
regulation, data protection, media and competition are large and
complex yet essential to understand the full picture of the regulatory
dynamics behind the algorithmic public opinion. In this Section, I
identify and present fundamental policy challenges and developments
that are particularly salient for the algorithmic public opinion. Of
course, I do not indend to provide any comprehensive guide to the
regulation of systemic issues. However, I aim to introduce the most
significant debates and current proposals around systemic issues
which are intended as complex regulatory issues that substantially
(and sometimes indirectly) influence the conditions for a healthy and
sustainable algorithmic public opinion ecosystem.

Algorithmic Regulation and Accountability

Algorithmic regulation – including issues related to the use of AI

and big data – has risen up the public agenda in recent years, with
a range of reports issued by international agencies, government
departments, legislative commitees, think-tanks and academic bodies.
Generally, attention has been given to a variety of issues such as raising
awareness on the functioning and (unintended) consequences of
algorithms (e.g. Algo:aware, 2018), the risks these entail for democracy
(e.g. AlgorithmWatch, 2020) and for discrimination (e.g. Zuiderveen
Borgesius, 2018) and, eventually, how to design algorithmic systems
responsibly (e.g. Alan Turing Institute, 2019), use audit methods in
social media (Ada Lovelace Institute, 2021) and assess the impact of
algorithms (e.g. AI Now Institute, 2018). Similarly, an intertwined strain
of analysis is concerned with AI more broadly; from ethics guidelines
(for example, HLEG AI, 2019) and human rights perspectives (e.g. EU
Agency for Fundamental Rights, 2020; Fjeld et al., 2020) to issues more
relevant for the algorithmic public opinion such as how AI-driven tools
in the media affect freedom of expression (Helberger et al., 2019).

Algorithmic regulation now features in so many different contexts

that it must engage a variety of legal rules beyond the data protection
sphere: administrative law, criminal law, intellectual property
law, contract law, and competition law are obvious cases in point.
There are also a large number of initiatives aimed at promoting
responsible algorithmic decision-making, including working groups
and committees, policy and technical tools, standardisation efforts
and, finally, codes of conduct, ethical principles and frameworks
(Algo:aware, 2018). Of course, it is unlikely that a single policy solution
or approach will deal with all, or even most of those challenges
discussed above. In order to address them, and to manage the tradeoffs

The Algorithmic Public Opinion: a Policy Overview 32

 that arise, a layered variety of approaches are likely to be required.
This section will therefore lay the ground for the subsequent analysis
of specific issues, in particular recommender systems and political
microtargeting, by introducing the main concepts and debates.
Algorithmic regulation is an emergent field of governance. Civil society
and industry have already begun to develop initiatives and design
technical tools to address some the issues raised throughout this
section. In particular, issues and questions around the benchmarks of
accuracy, fairness, accuracy and accountability are fundamental for the
development of policy toolboxes. During the 2010s the everincreasing
omnipresence of AI systems has been accompanied by a growing
importance attached to the ideals of fairness, accountability and
transparency in relation to algorithmic decision making (so-called
FAT); in a nutshell, fairness means the absence of systematic bias
and disadvantage towards particular demographics or social groups;
accountability is a tool that is supposed to contribute to fairness: it
has to be accounted for the algorithmic activities and accepted the
responsibility for the resulting (unfair) outcomes; transparency has to
do with visibility and insight into the system. In addition, there are often
discussed the principles of explainability – how AI outcomes can be
understood by humans (also called Explainable AI or XAI, see e.g. Adadi
and Berrada, 2018) and accountability – which refers to “a relationship
between an actor and a forum, in which the actor has an obligation to
explain and to justify his or her conduct, the forum can pose questions
and pass judgement, and the actor may face consequences. Thus an
‘actor’ (be they an individual, a group, or an organization) is required
to explain their actions before a particular audience, the ‘forum’ (24, p.
447) (for a literature review see Wieringa, 2020). There are a number
of techniques and approaches to verify the integrity of algorithmic
systems.

Algorithmic Impact Assessments. Principles are employed in

algorithmic impact assessments. An impact assessment can be defined
as “the process of identifying the future consequences of current
or proposed action.”10 By requiring an entity to conduct an internal
inspection, impact assessments urge coders and designers to conduct
a deeper form of analysis, carefully investigating plausible areas of
bias, error, and uncertainty, as well as implementing the necessary
steps to correct them. Of course, different jurisdictions have different
impactassessment schemes in place, and each has its own specificities
and objectives. Diakopoulos et al.11 elaborated the Principles for
Accountable Algorithms and a Social Impact Statement for Algorithms.
These are responsibility, explainability, accuracy, auditability, fairness.
Generally, however, impact assessments may provide only limited
transparency and allow only limited room for public review, for example
for automated content moderation (see Nahmias and Perel, 2021).

Human Rights and Risk-based Approches. Many of the issues raised

10 See the definition employed by the Inter- against principled-based governance can also be taken care of by
national Association for Impact Assessment
https://www.iaia.org
applying a human rights lens to algorithmic systems. One of the
11 https://www.fatml.org/resources/princi- earliest applications of the human rights framework to the topic of
ples-for-accountable-algorithms. AI was the Toronto Declaration - Protecting the right to equality and

The Algorithmic Public Opinion: a Policy Overview 33

 nondiscrimination in machine learning systems (2018). Since then
there has been an ever increasing amount of work in this area, with
academics, civil society organisations and international bodies all
publishing work on human rights and AI. Whereas voluntary ethics
guidelines leave large scope for companies to interpret what different
principles mean, the international human rights framework has
established mechanisms for resolving such ambiguities, and enforcing
compliance, even if that hasn’t always been without issues. Another
approach in this context is a risk-based one. These involve determining
the scale or scope of risks related to a concrete situation and a
recognised threat. This approach is useful in technical environments
where companies have to evaluate their own operational risks.

Algorithmic Auditing. Auditing techniques are a key part of a

regulatory inspection process to ensure accountability. Algorithmic
auditing enables interested third parties to probe, understand, and
review the behavior of the algorithm through disclosure of information
that enables monitoring, checking, or criticism. In the contex of
social media platforms, auditing can be undertaken using a variety of
techniques such as:

1. ‘Code audits’ (when auditors have direct access to the codebase of

the underlying system);
2. ‘User survey’ (when auditors conduct a survey and/or perform user
interviews to gather descriptive data of user experience on the
platform);
3. ‘Scraping audit’ (when auditors collect data directly from a plat-
form, typically by writing code to automatically click or scroll
through a webpage to collect data of interest);
4. ‘API audit’ (when auditors access data through a programmatic in-
terface provided by the platform that allows them to write comput-
er programs to send and receive information to/from a platform);
5. ‘Sock-puppet audit’ (when auditors use computer programs to im-
personate users on the platform and the data generated is record-
ed and analysed);
6. ‘Crowd-sourced audit’ (when real users are used to collect infor-
mation from the platform during use – either by manually report-
ing their experience, or through automated means like a browser
extension) (see Ada Lovelace, 2021).

Regulators need capacity, resources and skills to conduct these audits.

In addition, civil-society and academic actors should be enabled to
conduct these audits.

Algorithmic Debiasing. Furthermore, there is also the opportunity

for debiasing algorithmic systems, especially in content moderation
and search engines. Debiasing generally refers to “the application
of select methods to address bias by achieving certain forms of
statistical parity (EDRi, 2021, p.23).” As I have shown in Chapter
2, search engines ranking can indeed be biased. Kay et al. (2015)

The Algorithmic Public Opinion: a Policy Overview 34

 show, for instance, undesired biases in the Google image search
functionality when querying occupation-related images, with systematic
underrepresentation of women and stereotype exaggeration. In cases
like this, it is possible to ‘debias’ the algorithmic system. However, over-
relying on algorithmic debiasing can represent a narrow approach that
squeezes complex socio-technical problems into the domain of design
and thus into the hands of technology companies. Effective solutions
would actually require bold regulations that target the root of power
imbalances inherent to the pervasive deployment of AI driven systems.

Competition and Economic Incentives

For many reasons, digital markets – in particular search engines and

social media markets – tend to be highly concentrated with fundamental
entry barriers for potential competitors (Scott Morton et al., 2019).

These are in part due to certain characteristics of digital technology

(i.e. network effects), but in part also due to behaviors of market
participants-consumers create entry barriers with their behavioral
biases and companies by engaging in various activities. The resulting
concentrated market structures do not serve consumers.

Level the Playing Field. A fundamental policy approach to mitigate

many of the challenges and concerns coming from the algorithmic
public opinion is arguably levelling the “playing field” of information
intermediaries online, particularly resizing mainstream platforms and
supporting emerging innovative companies. We might indeed expect
the marketplace to self- correct (to some extent) and for companies that
offer tools that are, for example, more privacy and autonomy-preserving
to gain a competitive advantage. Such policy could be done by breaking
big tech companies (for instance Whatsapp from Facebook, and Youtube
from Google) and, at the same time, preventing them from acquiring
potential competitors. So far, users had only an illusory choice: take it or
leave a bunch of platforms. The problem is indeed systemic. Consider
how in the past 20 y ears the GAFAM (Google, Apple, Facebook, Amazon,
Microsoft) collectively bought 1,000 firms, and 97% of these transactions
have not been vetted by anyone, very few challenged, and zero blocked
in the US and around the world as well (Cabral, 2021). The result is
that these companies became ever more powerful while we missed
the impact of competition in so many different ways. It is also true,
however, that to the extent that competition regulation would fragment
intermediary markets and disperse market power, it might make harder,
not easier, for online content harms to be addressed (Bunting, 2018).

Business Model. Unless the successfulness of the business model

based on advertisements is completely undermined, we can’t expect
the market to easily correct itself. Many scholars argued that this
business model inevitably built on engagement and popularity lead
to divisive, emotional content, because that is what algorithms tend
to favor. So, what could be the alternative to the attention economy
business model? The most common and viable potential solution is
changing to a subscription-based model. This would allow businesses

The Algorithmic Public Opinion: a Policy Overview 35

 to own the relationship with consumers. Yet, this would reduce access
to fundamental, ordinary services and information for those who can’t,
or won’t, subscribe. This potential exclusion is a serious challenge to
buildup alternative and sustainable social media business models.
Perhaps a more convincing solution may be found focusing on the
economic incentives, namely ads.

Tax, Ban or Limit Online Ads. The AD industry should help sellers to
understand customers and deliver them marketing messages that are
more relevant, consistent and effective. In general, it is argued that this
model helps to fund the press and other channels of expression. Critics,
instead, argue that this model incentivizes sensationalistic journalism,
clickbait and, overall, negatively affects the quality of the press (Kingaby
and Kaltheuner, 2020). More often, misinformation and conspiracy
theories are the product of this business model, not an accident. For
many reasons – among them adblockers (plug-ins to block ads), outof-
sights ads (ads that are rarely seen by users) and click-farms (fake ads
views) – it can be even radically questioned the effectiveness (and thus
returns) of most online ads (Neumann et al., 2020). Not only there is
little evidence that constant tracking leads to more relevant ads, but
a recent study showed how targeted advertising accounted for only
a percent average increase in revenue. Of course, Google has argued
that publishers would lose half their revenue or more if they stop using
personalized advertising. Actually, online ads seem so over-valued
that might even represent the next financial bubble (Hwang, 2020). A
radical solution that critics advocate is to ban micro-targeted ads and
to opt instead for contextual ads, not only for their unintended and
undesirable consequences but also because they appear to be limitedly
effective. In addition or in alternati ve, a compelling policy approach
is proposed by the Nobel prize for economics Paul Romer (2021) who
argues to enact a progressive, sufficiently aggressive tax on revenue
from digital advertising. This could make the subscription model more
attractive or, more simply, to make it more attractive for a large firm to
create independent new ventures, and less attractive for it to grow via
acquisitions. To date, there is no policy agenda supporting these latter
solutions. In EU, however, there are currently more than 20 proposals
for behavioral ads reform being considered by legislators working on the
Digital Services Act and Digital Markets Act.

Sustain Alternative Social Media. Mainstream platforms’ business

model could also be challenged by rival ideas and competitors. These
competitors must stand a chance, not only thanks to active antitrust
enforcement but also by seriously taxing large media (and other)
corporations and channeling the resulting income into alternative
noncommercial social media (Fuchs & Marisol, 2015). Indeed, so far the
history of alternative media is a history of enormous challenges, mainly
because hearing alternative voices is ultimately a matter of money and
political resources to afford visibility. To some extent, this challenge
12 https://www.ngi.eu/.
is considered in EU. A paradigmatic example is the European project
Next Generation Internet (NGI)12. This has funded, among many other
13 https://framablog.org/2021/11/30/peertube-
v4-more-power-to-help-you-present-yourvid-
projects, PeerTube – a free, libre and federated video platform – with
eos/. 50.000 Euros.13 Obviously, it might be contested that the current funding

The Algorithmic Public Opinion: a Policy Overview 36

 is insufficient to scale up and compete with mainstream platforms.

Strenghten Interoperability. Another promising policy solution is to

strenghten interoperability14. Digital interoperability enables apps,
digital services and devices to work together, even if they are made by
different providers. Well-known examples of interoperability include
e-mail, and telephone voice and messaging services — you can send
an e-mail or text message or call anyone else, regardless of the service
providers, apps or devices you use. In contrast, mainstream social media
services like Facebook or Twitter as well as many other services, tend to
only support interactions within their own platforms (e.g. a Facebook
user cannot follow someone else’s Twitter feed or a Telegram user is
blocked from joining a WhatsApp group). This requires consumers
to use multiple applications and devices that are incompatible with
each other. Interoperability is one of the basic principles on which the
internet was built. By adopting open technical standards, it would be
possible to break down ‘walled gardens’ controlled by a single company.
Interoperability of digital services can stimulate innovation by allowing
new operators to enter markets such as social media and messaging
services. This in turn creates an incentive for all operators to innovate
and provide new features. Currently, mainstream platforms can rely
on their network effects without having to compete on the merits of
their products and services. Mandatory interoperability for the largest
digital platforms can also foster the creation of whole new digital
markets where startups build digital services on top of incumbent
platforms. That way, consumers can get access to better AI-driven
content moderation algorithms that run on top of a user’s Twitter feed or
Facebook timeline. In the same vein, new apps could replace YouTube’s
recommender algorithm that is known to promote extremist video
content even to people who weren’t looking for it.

Data Protection and Privacy

Algorithmic systems have the potential to transform seemingly

non-sensitive data into sensitive data about individuals. At times, such
transformations can create the possibility for discrimination against
individuals and groups, or simply violate expectations and create data
flows and knowledge that individuals may find inappropriate. A range
of issues surround the transformation of data from sensors or online
behaviour into sensitive data that concerns an individual’s health,
wellbeing or mental state.

Data protection is, of course, essential for privacy. It is fundamental

not only to avoid possible manipulations of the algorithmic public
opinion but also for individual and collective human rights. Of
course in this section I do not intend to raise all the important issues
concerning privacy and data security that relate to the algorithmic
public opinion. There are several other efforts underway that aim to be
more comprehensive than this one, for example on the data ecosystem
and governance (e.g. World Bank, 2021), the debate on data sovereignty
14 For more information on interoperability see
(e.g. De La Chapelle and Porciuncula, 2021), data ownership and the
https://interoperability.news/. future of data (e.g. Decode, 2019; Mills, 2020; Ada Lovelace Institute,

The Algorithmic Public Opinion: a Policy Overview 37

 2021). I hope, however, to shed lights on a number of open technical and
policy questions and debates such as: how to overcome the limitations
of platforms’ informed consent? When dark patterns are manipulative
should they be outlawed? And what new empowering rights need data
consumers in this context? How to build a new paradigm for data
ownership in this context?

Protect Group Privacy. Due to a set of externalities and information

leakages involved in data markets, privacy could be actually recognized
as an “aggregate public good” prone to market failure. Recognizing this
should convince us that government intervention is both beneficial and
necessary for its protection. This is even more relevant considering the
flaws of privacy self-management and emerging algorithmic techniques
that might threaten the privacy of groups. In fact, algorithmic systems
which measure, count, and profile groups of individuals create
knowledge that is not (only) private to an individual, but which reveals
something about a group of individuals. This is framed as the protection
of group privacy. Since long privacy researchers proposed a contextual
and relational understanding of privacy, mainly refferred to as relational
privacy. The main problem is that algorithms can create ad hoc and
temporary groups to which none of the existing interpretations of
privacy can be applied. The implication of the technical issues related to
group privacy is that our legal, philosophical, and analytic attention to
the individual may need to be adjusted, and possibly extended (Taylor et
al. 2016). The fact that the individual is no longer central, but incidental
to these types of processes, challenges the very foundations of most
Western legal, ethical, and social practices and theories related to
privacy.

Users’ Data as a Public Resource? As regulatory framework, Napoli

(2019) proposes to treat users’ data as a public resource. The central
premise is that, whatever the exact nature of one’s individual property
rights in one’s user data may be, when these data are aggregated across
millions of users, their fundamental character changes in such a way
that they are best conceptualized as a public resource. Certainly, it is in
this massive aggregation that the economic value of user data emerges.
Therefore, if policymakers would treat aggregate user data as a public
resource akin to broadcast spectrum, that framework would provide
what may be the most constitutionally robust rationale for imposing
public interest obligations upon those social media platforms that rely
upon the aggregation and monetization of user data. This approach is
legally grounded as well as promising for it can allow researchers and
civil society to access and analyze social big data. In EU, the Art. 31 of
the proposed Digital Services Act would give researchers with academic
affiliations access to platform data for public interest research (see
Leerssen, 2021).

Beyond Individual Consent? The individualistic approach of privacy

self-management usually relies on informed consent, but this seems
not be an optimal solution because it leads to uncertainty and context
dependence. People in fact cannot be counted on to navigate the
complex trade-offs involving privacy self-management (Acquisti et al.,

The Algorithmic Public Opinion: a Policy Overview 38

 2015). Most people in fact neither read nor understand online privacy
policies. These have also two inherently contradictory goals: to be
understandable to consumers – which requires simplicity and brevity –
and say something meaningful about how data is processed – which is
complicated and requires a lot of details. As well known, it would take
an enormous amount of time to read all the conditions of the websites
we visit – 201 hours on average per year. Moreover, dark patterns are
often employed during terms of conditions and privacy updates to
nudge consumers toward options that benefit company profitability but
may not reflect consumers’ actual preferences or expectations (Moen
et al., 2018). Even worse, the least educated seem most likely to be
manipulated successfully (Luguri and Strahilevitz, 2021).

As Strahilevitz et al. 2021 argue, a compelling solution to this model

is that the content of contractual default provisions would depend
on the articulated preferences of ordinary consumers as measured
by scientifically rigorous survey instruments. In privacy and security
settings there are in fact many instances in which it is appropriate for
the law to use “consumertarian” default rules – i.e., the legal defaults
preferred or expected by a majority of consumers. This would arguably
alling users’ preferences to platforms’ terms and conditions. Of course,
such an approach is not devoid of implementation challenges. More
generally, nonetheless, the development of informed consent with
respect to big data use is highly debated and other proposals are indeed
discussed (see Andreotta et al., 2021).

A Ban to Dark Patterns? Many of the inherent problems with dark

patterns have implications for information privacy. Dark patterns are
indeed often used to direct users toward outcomes that involve greater
data collection and processing. Additionally, the proliferation of data-
driven computational methods allows firms to identify vulnerabilities
of users and to target specific users with these vulnerabilities. While
dark patterns come in a variety of different forms, their central unifying
feature is that they are manipulative, rather than persuasive. More
specifically, the design choices inherent in dark patterns push users
towards specific actions without valid appeals to emotion or reason.
The line between manipulation and persuasion, however, is sometimes
difficult to draw, not only ethically but also legally. Strahilevitz et al.
2021 propose a framework that could allow legislators, regulators, and
courts to define the category of manipulations warranting legal action
in a way that is workable and defensible on both economic and moral
grounds.

Editorial Obligations and Intermediary Liability

Considering the special position of media and the preference for

self-regulation in this area, for some time it has been considered
unlikely that the normative principles in the media context would
directly translate into legal obligations. Recent debate on the public
responsibility of social media platforms pivots on the question of
whether or not platforms can be held accountable for the content shared
through them, legally and morally. Platforms indeed still enjoy the same

The Algorithmic Public Opinion: a Policy Overview 39

 status as Internet Service Providers (ISPs) under US (Communications
Decency Act, 1996) and EU law (E-COMMERCE Directive 2000/31/EC).
The growth of information intermediaries, and their significant control
over distribution of certain kinds of content, has in fact reinvigorated
public debate about the appropriate balance between competing rights
to free expression and protection from harmful or illegal content.
Today there is an increasing consensus - especially in the European
Union - that due to market concentration new legal approaches and
new governance models have become necessary. New definitions seem
to be needed to address the role of the information intermediaries and
distinguish them from ISPs. Mainstream platforms are difficult to frame,
and their consequences are equally difficult to govern and regulate.

The actual role and capacities of social media’s platforms to prevent

certain undesirable outcomes or to contribute to their realization is still
debated. Of course, platforms fundamentally shape user activity, yet
they do not determine this activity. Many of the problems with media
pluralism and diversity are, to a large extent, user-driven. For similar
reasons, at least part of the solution to potential public policy challenges
lies with the users. The current focus in law on allocating responsibility
to one central party – editor, data controller, or the supplier of a service
– is primarily due to the fact that this central actor is the source of
potential risk or harm, or the controller of a resource that can give rise
to legal responsibilities. Yet, multiple actors are effectively responsible.

From a legal point of view, this discussion is grounded in the

host-editor, namely either social media qualify as hosts, with the
consequence that they fall under the European e-Commerce regime,
or they are categorized as editors, having full legal responsibility for
what is shared through their platforms. As many scholars have argued,
the legally enshrined conceptual framing of a “platform” that merely
hosts content, but should not be held legally liable for it, became a
strategic and powerful enabler for the rise of today’s digital giants.
Ample scholarship has since shown that the framing of technology
companies as mere “hosts” or “intermediaries” or “platforms” elides the
ways in which these companies set norms around content or speech ,
algorithmically recommend content, and assume a host of functions that
combine features of publishers, media companies, telecommunications
providers, and other firms (Napoli and Caplan, 2017).

Yet, there are three specific problems with imposing editorial

obligations on intermediaries:

1. The lack of accountability. Editorial obligations tend to increase

intermediaries’ power over content markets, without necessarily
increasing the transparency or accountability of its use. Large vol-
umes of material may be removed or blocked, with little visibility
of the true social benefits and costs.
2. The impact on competition. Making intermediaries responsible for
the content they host may have a disproportionate impact on new
entrants for whom the cost of preventing content infringements

The Algorithmic Public Opinion: a Policy Overview 40

 would represent a substantial burden. Most intermediaries process
a vast volume of content, much greater than any traditional pub-
lisher, meaning that the costs of proactive monitoring of content
are also much greater. Yet, regulation can refer to specific require-
ments that platforms should have to avoid to burden emerging
platforms.
3. Imposing editorial obligations on intermediaries may have limited
effect. Harmful content can easily flow from more regulated to less
regulated environments, where it may be less visible and less sus-
ceptible to responsible intermediary activity. For example, aften
the removal of Trump on Facebook and Twitter millions of users
moved to the right-wing platform Gab.

Editorial obligations skew intermediary incentives towards those

interests which are protected by takedown; the stronger the incentives
on intermediaries to remove content rapidly, the greater the likelihood
of legal content being inadvertently blocked. These cases raise questions
of fundamental rights, and how rights that may be in tension are to be
balanced. The goal of blocking illegal content must be reconciled with
the risk of inadvertent denial of access to legal content. But editorial
obligations and liabilities create no incentive for intermediaries to
consider a ‘fair’ or ‘just’ balance, only to secure the commercially
optimal outcome, which will skew towards content takedown in
proportion to the size of the sanction for distributing illegal content.
The existing content regulatory toolkit is wholly unsuited to the task.
New ways of regulating intermediaries – which reconcile their market
governance with protection and balancing of rights – are needed.

Specific Issues

To complement systemic issues, in this policy review I also focus

on a number of more specific issues I have identified and that are
paramount, namely recommender systems, content moderation, social
bots, political microtargeting and, finally, disinformation. All these
are certainly intertwined with the above systemic issues. To begin,
the analysis of recommender systems is stricly related, in particular,
to the governance of algorithms, and it is fundamental to develop
personalized filtering that are fair and accountable. Similarly, political
microtargeting has to do with algorithmically personalized political
advertisement. Then, content moderation is equally relevant in the
content management of social media platforms, while social bots are
a concerning weapons of online manipulation, often disrupting the
marketplace of ideas. Finally, disinformation is a more systemic issue
that includes all of the previous issues and, in fact, has received more
attention from public opinion and policy-makers.

Recommender Systems

Recommender systems (RSs) are probably the most relevant

algorithmic system in the context of the algorithmic public opinion.
In particular, RSs generate serious concerns on hate speech,

The Algorithmic Public Opinion: a Policy Overview 41

 disinformation and conspiracy theories, but also on monopolisation
and platform power. Legal responses to these problems are not
straightforward given the various stakeholders and the amount of
information that platforms typically deal with. Any form of restriction
on recommendations is difficult to automate, culturally contextual, and
potentially sensitive. Also, regulation focusing on the transmission or
hosting of content itself brings freedom of expression concerns. Yet, as
Cobbe and Singh (2019) argue, the same risks do not necessarily arise
from regulating the further dissemination of content by platforms.
While the fundamental right of freedom of expression should be
respected as far as possible, individuals do not have a fundamental right
to have their speech disseminated or amplified by platforms in this
way (as the adagio goes “freedom of speech is not freedom of reach”).
By focusing on recommending, rather than on the transmission or
hosting of content itself, regulation can largely sidestep these freedom
of expression problems and focus on the use of technical systems by
private corporations to pursue their own business goals.

For recommender systems to be ‘responsible’, Cobbe and Singh

(2019) outlined a number of principles for the service providers who
use these systems, particularly (open) RSs systems15. In short, they
must be lawful and service providers should be prohibited from doing
it where they violate these principles. Indeed, if service providers can’t
use RSs responsibly then they shouldn’t be permitted to do it at all.
Similarly, service providers should have conditional liability. No liability
protection would therefore be available when undertaking a prohibited
practice. Service providers should indeed have a responsibility to not
recommend certain ‘potentially problematic content’. This should
establish a responsibility to not promote certain kinds of content (for
example, white supremacism, health disinformation, anti-Semitic
conspiracy theories, pro-suicide or self-harm content, content
promoting eating disorders, and so on). Any potential liability would
therefore actually result from undertaking a prohibited practice rather
than from recommending certain kinds of potentially problematic but
15 To be more specific, in fact, the following
principles refer to ‘open recommenders’ which
lawful content.
are those that provides recommendations from
a pool of content which is primarily user-gen-
erated or submitted, brought in automatically
from various sources, or otherwise aggregated Individual Control. First of all, RSs should be opt-in, meaning
in some way without being specifically selected users should be able to exercise a minimum level of control over
by the platform. Examples include Google, You-
Tube, Facebook, Reddit, Instagram and Amazon. recommending, and opting-out again should be easy. Offering control
These differentiate from ‘curated recommend- to users is a good idea if, and only if, those who do not exercise it do not
ing’ in which the system recommends from a
pool of content which is curated, approved, or
end up being treated less favourably than those who do. To this end,
otherwise chosen by the platform rather than it would be desirable for RSs to be available to users only on an opt-
provided directly by users or advertisers or
in basis. Users who choose to receive recommendations should, at a
automatically brought in from elsewhere and
do not typically include user-generated content minimum, be able to: exclude certain content from recommendations,
without some kind of editorial process (Netflix is exclude certain sources of content from recommendations, exclude
a popular example of a curated system). Finally,
they are distinguished from ‘closed recommend- certain of their behaviours or interests from the process of determining
ing’ where the content to be recommended is what should be recommended to them, and to easily and freely opt back
generated by the platform itself or the organ-
isation which operates that site. For example, out of recommendations entirely. Furthermore, users could be enabled
where a news organisation provides a person- to choose between different RSs, including some from third parties.
alised feed of stories and articles to its users, all
of which are produced or commissioned by the
organisation itself. User-facing disclosures. Similarly, user-facing disclosures aim to

The Algorithmic Public Opinion: a Policy Overview 42

 channel information towards individual users in order to empower them
in relationship to RSs (e.g. Facebook’s ‘Why Am I Seeing This’ feature).
The aim of such transparency is to inform users about their available
options so as to help them form their own preferences, appealing to
values such as individual autonomy, agency and trust.

Public Disclosures. Another significant transparency requirement is

that service providers should be required to keep records and make
information about recommendations available to help inform users and
facilitate oversight (i.e. public disclosures). The constantly changing
nature of social media and other online services makes it difficult to
identify and track problems over time. Service providers should, at a
minimum, be required to keep logs of recommended content (both
for personalisation and for behavioural targeting) so that they can
be reviewed by users and by oversight bodies (for a discussion see
Leerssen, 2020). Of course, provided that these should have privacy-by-
design and a trustless design that pre-empts abuse by malicious actors,
public records could be instrumental for purposes of real-time, high-
level monitoring by media watchdogs such as journalists, activists,
and NGOs. These may not suffice to conclusively demonstrate bias or
discrimination in RSs, but at a minimum they could offer a starting point
for such investigations and serve as a first-warning system for more
targeted efforts.

New Oversight Authority. Government oversight could appoint a public

entity to monitor RSs for compliance with publicly regulated standards.
This endeavior faces many significant challenges, both practical and
principled. To begin, government authorities are capacity-constrained,
particularly with regard to the technical expertise required to perform
complex algorithmic auditing. This is especially true for horizontal
agencies such as competition and data protection authorities, for whom
RSs risk being overshadowed and overlooked. Sectoral proposals,
instead, would in many cases require the creation of entirely new
oversight bodies. Government auditing powers also raise issues of
‘second-order accountability’: is the governance system itself sufficiently
open to outside scrutiny? If government determinations rely on
privileged access to confidential data, which is not accessible to broader
publics, it may be difficult for citizens to fact-check and second-guess
government policy in this space. Without broader forms of second-
order transparency and accountability, the legitimacy of a technocratic,
command-and-control approach in such a politically sensitive, value-
laden context can therefore be called into question.

Access to Data for Research. Partnerships with academia and civil

society would better enable these stakeholders to research and critique
RSs. Yet, these are often met with skepticism for several reasons.
Above all, creating meaningful transparency arguably runs counter to
platforms’ incentives: they have a commercial interest in monetizing
traffic data and insights, and thus in keeping this information
exclusive, and a political interest in avoiding negative publicity. Besides
independent surveying, one of the most important sources of data
regarding recommender systems has been their public APIs, through

The Algorithmic Public Opinion: a Policy Overview 43

 which outside researchers can download platform data in bulk. But
these have come under significant pressure over the past years. In the
EU, it is finally discussed the regulation of researcher access to platform
data. Article 31 of the proposed Digital Services Act (DSA) on “Data
Access and Scrutiny” is the first legislative framework for researcher
access to platform data. It has been welcomed with a mix of hopes and
criticisms (Leerssen, 2021).

Other Restrictions. There should be also specific restrictions on service

providers’ ability to use recommending to influence markets through
RSs. Service providers should be explicitly prohibited from unduly
recommending their own products and services ahead of those offered
by others. These prohibitions would not only complement and refine
the existing data protection principle of purpose limitation, but would
go some way towards addressing competition issues arising from the
dominance of platforms and their use of personal data, particularly
where leveraging their dominant position in one market to gain a
competitive advantage in another. Beyond these general principles
for ‘responsible RSs’, other applicable legal frameworks must also
be considered. Data protection, in particular, is fundamental in this
context, given the extensive behavioural tracking and processing of
personal data which underpins RSs. Similarly, service providers cannot
ignore their responsibilities under equality and non-discrimination law
or, indeed, any other applicable regime.

Content Moderation

Content moderation not only involves algorithms but also users

in conducting review processes, which requires them to set content
standards that are able to be easily encoded, interpreted and applied
consistently across widely varying national jurisdictions. This would
suggest that platforms engage in moderation similarly to the editorial
and governance processes undertaken by legacy media companies
in content regulation. Yet, the user led nature of content generation,
the scale of creation, the inclusion of community reporting and the
decentralization of decision-making across time zones and cultural
contexts makes these processes more diverse, complex and demanding
to negotiate than traditional, professionally oriented editorial decisions.

The massive user generated flows of digital platforms today have

been built on the premise that content can be post moderated –
although AI-oriented automation will eventually enable high degrees
of algorithmic pre-moderation. However, there is a large disjuncture,
according to Gillespie (2018), between the ‘data scale’ of faceless,
consistency oriented, automated regulation and the ‘human scale’ of
localized, culturally bound interactions, which suggests that machine
controls will never be a complete solution to the challenge of classifying
endless content flows. Addressing the problem of national differences
in content regulation and cultural expectations of publishers is also key
to the future of platform governance. At the same time, there is a myth
that these systems are completely automated: there is always a decision-
making process behind categorising content, behaviours and people as

The Algorithmic Public Opinion: a Policy Overview 44

 deviant, but so far these have been hidden and unaccountable.

Furthermore, public opinion have mostly focused on the latest

controversies and the biggest players. Major platforms are in fact
enormous and their policies affect billions of users. Their size
makes them desirable venues for malicous actors. Their policies and
techniques set a standard for how content moderation works on other
platforms. Yet, the largest, US-based platforms do not provide a reliable
guide for the entire social media ecology. There are many kinds of
social media platforms that configure content moderation differently.
Moderation also happens on sites and services different to mainstream
social media: on comment threads and discussion forums, in multi-
player game worlds, in app stores, on dating sites, and on the many other
services. These also differ in ways that affect how content moderation
works: by size, reach, and language, but also by technical design, genre,
corporate ethos, business model, and stated purpose.

The content moderation debate should expand beyond treating

platforms as primarily venues for public speech, or as silos that exist
in isolation from one another (Gillespie et al., 2020). Instead, we might
think of them as a web of private infrastructures that we traverse in our
digitally mediated lives. They are indeed also marketplaces, payment
systems, advertisers, gaming sites, and media distributors. Rather than
thinking about content moderation in terms of its effects on speech
alone, we should instead consider the consequences that moderation
can have on communities, by influencing the access to platforms that
are increasingly central to our ability to work, live, and socialize.

There are also a number of critical areas to improve moderation

and minimize its risks. Innovations in automated content moderation
have focused overwhelmingly on identification techniques such as
detect pornography, harassment, or hate speech. Not only are there
problems with these ambitions but automated content analysis tools
have shadowed other possible uses of algorithms to support content
moderation. Research should also prioritise tools that might support
human moderators, community managers, and individual users to
eventually make more informed decisions and data-scientific techniques
might also help users and community managers better grasp how
differently other communities experience similar content or behaviour.

Another underdiscussed but significant issue is the recording

of the content that has been removed. In particular, social media
documentation of human rights violations is critical for justice and
accountability efforts, and in some cases it serves as collective memory.
Videos and text posted online are living histories for some diaspora
communities, and sometimes this documentation might offer the only
evidence that a crime has been committed. Yet in too many cases,
social media content moderation policies around extremism lead to the
deletion of vital documentation. Restoring wrongfully deleted content
is nearly impossible if the person who posted the content is not alive, is
arrested, or does not have access to email, all common issues in conflict
zones.

The Algorithmic Public Opinion: a Policy Overview 45

 Nonetheless, as much as platform content moderation could improve,
it may also be a perennially impossible task to do in such a way that
no one encounters harm, friction, or restriction. Users of social media
may have unreasonably high hopes for what their experience should be,
largely because of the endless promises made by social media platforms
that it would be so. We need to educate and adjust the expectations of
users, to both understand what a difficult and vital process this is, to
demand it be transparent and accountable, to recognise how they are
implicated in it, and to prod their sense of agency and ownership of
these sometimes unavoidable dilemmas.

Social Bots

Gorwa and Guilbeault (2019) argue that multiple forms of ambiguity

are responsible for much of the complexity underlying contemporary
bot-related policy, and that before successful policy interventions can be
formulated, a more comprehensive understanding of bots—especially
how they are defined and measured—will be needed.

Any initiatives suggested by policymakers and informed by research

will have to deal with several pressing challenges: the conceptual
ambiguity (highlighted in the Chapter 1), poor measurement and
data access, lack of clarity about who exactly is responsible, and the
overarching challenge of business incentives that are not predisposed
toward resolving the aforementioned issues.

Measurement and Data Access. Bot detection is very difficult.

Researchers are in fact unable to fully represent the scale of the
current issue by relying solely on data provided through public APIs.
Notoriously, they cannot study bots on Facebook and virtually all studies
of bot activity have taken place on Twitter which already questions
whether their APIs provide a fair account of content on the platform.
Even the social media companies themselves find bot detection a
challenge, partially because of the massive scale on which they function.
Tracking the thousands of bot accounts created every day, when
maintaining a totally open API, is virtually impossible. Taking this a step
further by trying to link malicious activity to a specific actor (e.g., groups
linked to a foreign government) is even more difficult, as Internet
Protocol (IP) addresses and other indicators can be easily manipulated.

Even the most advanced current bot detection methods hinge on the
successful identification of bot accounts by human beings. The problem
is indeed that humans are not particularly good at identifying bot
accounts. Researchers can never be 100 percent certain that an account
is truly a bot, posing a challenge for machine learning models that use
human-labeled training data. The precision and recall of academic bot
detection methods, while constantly improving is still seriously limited.
Of course, less is known about the detection methods deployed by the
private sector and contracted by government agencies, but one can
assume that they suffer from the same issues. Just like researchers,
governments have data access challenges. This pose substantial
challenges to identify the scale of bot activity, epsecially during

The Algorithmic Public Opinion: a Policy Overview 46

 elections. The policy implications of these measurement challenges
become very apparent in the context of the recent debate over the role
of Russian propaganda during the 2016 U.S. Presidential election. To
understand the scope and scale of the problem, policymakers will need
more reliable indicators and better measurements than are currently
available.

Responsibility. A key, and unresolved challenge for policy is the

question of responsibility, and the interrelated questions of jurisdiction
and authority. To what extent should social media companies be held
responsible for the dealings of social bots? And who will hold these
companies to account? A whole spectrum of regulatory options under
this umbrella exist, with some being particularly troubling. For example,
some have argued that the answer to the “bot problem” is as simple as
implementing and enforcing strict “real-name” policies on Twitter—and
making these policies stricter for Facebook. The recent emergence
of bots into the public discourse has reopened age old debates about
anonymity and privacy online, now with the added challenge of
balancing the anonymity that can be abused by sockpuppets and
automated fake accounts, and the anonymity that empowers activists
and promotes free speech around the world. In a sense, technology
companies have already admitted at least some degree of responsibility
for the current political impact of the misinformation ecosystem, within
which bots play an important role. The matter is by no means settled,
and will play an important part in the deeper public and scholarly
conversation around key issues of platform responsibility, governance,
and accountability.

Contrasting Incentives. Underlying these challenges is a more

fundamental question about the business models and incentives of
social media companies. Business incentives are indeed critical in
shaping content policy—and therefore policies concerning automation—
for social media companies, slightly different incentives have yielded
differing policies on automation and content. Notably, Twitter’s core
concern has been to increase their traffic and to maintain as open a
platform as possible. Thus, it allows to use tools that automate their
activity, which can be useful: accounts run by media organizations, for
example, can automatically tweet every time a new article is published
but also fulfill many creative, productive functions. Facebook, instead,
has been battling invasive spam for years and has much tighter controls
over its API. As such, it appears that Facebook has comparatively much
lower numbers of automated users (both proportionally and absolutely)
than Twitter, but is concerned primarily with manually controlled
sockpuppet accounts, which can be set up by anyone and are difficult or
impossible to detect if they do not coordinate at scale or draw too much
attention. Thus, platform interests often clash with the preferences of
the academic research community and of the public. Academics strive
to open the black box and better understand the role that bots play in
public debate and information diffusion, while pushing for greater
transparency and more access to the relevant data, with little concern
for the business dealings of a social networking platform.

The Algorithmic Public Opinion: a Policy Overview 47

 There are no easy solutions to these challenges, given the complex
tradeoffs and differing stakeholder incentives at play. As a highly
political, topical, and important technology policy issue, the question
of political automation raises a number fundamental questions about
platform responsibility and governance that have yet to be fully
explored by scholars. Conceptual ambiguity can be reduced by diligent
scholarship, and researchers can work to improve detection models, but
responsibility may not be taken and business incentives will not shift on
their own. Despite mounting concern about digital influence operations
through social bots over social media, especially from foreign sources,
there have yet to be any governmental policy interventions developed
to more closely manage their political uses. An important measure has
been proposed by state legislators in California in April 2018. This was
requiring that all detected bot accounts were publicly labeled by social
media companies.

Political Micro-targeting

Political micro-targeting (or ‘behavioral advertising’, ‘political ads’ or

even ‘dark ads’) is defined as ‘creating finely honed messages targeted
at narrow categories of voters’ based on data analysis ‘garnered from
individuals’ demographic characteristics and consumer and lifestyle
habits’. It can be summarised as consisting of three steps: 1) collecting
personal data, 2) using those data to identify groups of people that are
likely susceptible to a certain message, and 3) sending tailored online
messages (Zuiderveen Borgesius et al., 2018). As such, unlike traditional
political advertising, micro-targeting not only affects the democratic
process, but it also affects people’s privacy and data protection rights.
Indeed, micro-targeting affects myriad other rights and duties, including
a political party’s and online platform’s right to impart information, a
voter’s right to receive information, and the government’s duty to ensure
free and fair elections.

The objectives of political micro-targeting can be manifold: to

persuade, inform, or mobilise, or rather to dissuade, confuse or
demobilise voters. It can conceivably be used in interesting ways to
be part of government communication and thereby unfold not only
as a feature of campaigning but also of governing. People can be
micro-targeted on the basis of all kinds of information (such as their
personality traits, their location, or the issues they care about). Hence,
any data can be valuable: from consumer data to browsing behaviour.
Such data can provide enough information to make inferences
about the susceptibilities of the target audiences. A micro-targeted
audience indeed receives a message tailored to one or several specific
characteristic(s). This characteristic is perceived by the political
advertiser as instrumental in making the audience member susceptible
to that tailored message. For example, when micro-targeting a party
could ignore the unlikely voters and tailor their messages to possible
voters' issue salience (or other characteristics). During the Brexit
referendum, the cross-party “Vote Leave” campaign commissioned well
1,433 customized adverts promoting a more or less explicit pro-Brexit
message (Reviglio & Agosti, 2020). A regular targeted message, instead,

The Algorithmic Public Opinion: a Policy Overview 48

 does not really consider matters of audience heterogeneity.

Micro-targeting originates from the United States, where there are

relatively loose data-protection. It seems likely that, when compared to
the US, Europe’s privacy rules hinder micro-targeting. Anyway, national
parliaments seem best placed to regulate political micro-targeting,
especially for what concerns electoral laws.

There are several concerns related to political microtargeted ads.

Despite changes that were introduced by most platforms, political
advertising continues to lack the transparency necessary to ensure fair
and democratic elections. Many changes introduced by the platforms
haven’t been rolled out globally, meaning countries with volatile political
contexts and fragile democracies risk being most vulnerable to election
interference. According to a 2019 study by Privacy International,
Facebook only required political advertisers to be authorised, or for
political ads to carry disclosures, in around 17% of countries around
the world. Google provides ‘heightened transparency’ for political ads
in 30 countries – around 15%. Furthermore, political microtargeting
techniques can also amplify the effects of deepfakes, but for a much
smaller subgroup than expected (Dobber et al., 2019). Deepfakes can
(see → Box 4 The Rise indeed poison the public debate by confusing people on what is real and
of Deepfakes) what is not. A number of counteractions might certainly mitigate the
challenges that political micro-targeting poses.

Conceptual clarity. Definitions of political ads vary widely. Thus, a

consensus is needed, at least in specific contexts.

More Research. We still need to understand how benefits of

microtargeting outweigh the risks. Therefore, research is needed,
not only to bring the above conceptual clarity but also to assess the
effectiveness of these techniques as well as help to develop ways to
scrutinize these delicate processes of public opinion formation. More
generally, it is fundamental both to reach more transparency and to
access relevant data.

More transparency. Necessary but not sufficient, there are a number

of potential transparency requirements: from expenditures of political
parties and from intermediaries’ ads, we need to set a regulatory
framework that makes ‘ads central repositories’ accountable. It is
possible to legislate so that all paid-for political adverts can be viewed
by the public. Yet, when it comes to such political ad libraries, there is a
lack of standardisation. In any case, these should be able to inform users
to see if:

1. A political advertiser microtargeting;

2. A true, verified name of the advertiser in the disclaimer about who
paid for the ad.
3. Eventually require political parties to disclose their campaign
finances broken down by media outlet.
4. How a platform has amplified the ad.

The Algorithmic Public Opinion: a Policy Overview 49

 5. Give an existing body the power to regulate political advertising
content or create a new one to do so.
6. require all factual claims used in political adverts to be inde-
pendently substantiated.

Limit political ads usage. Policy-makers could probably ban online

political micro-targeting, at least for a period leading up to elections
and during the run-up like referendum. Many countries have already
sector-specific rules for political advertising, which differ from country
to country. For decades, paid political advertising on television has been
completely banned during elections in many European democracies.
These political advertising bans aim to prevent the distortion of the
democratic process by financially powerful interests, and to ensure a
level playing field during elections. Member States’ rules for political
television advertising still differ widely from what is required for online
political advertising. Policy-makers could nsure that rules for online
political advertising match those for offline political advertising on
traditional media in the context of local, national or sovra-nationals (e.g.
EU) elections.

Limit psychometric profiles. Psychometric profiles could be labeled

a ‘special category’ of data. Citizens seem to react differently to affect-
based political ads based on their psychometric profile: introverted
people generated higher voting intentions when they were targeted
with a negative fear-based political ad, whereas extraverted citizens had
higher voting intentions after receiving a positive enthusiasm-based
political ad. This means that the processing of data revealing a person’s
psychometric profile is only allowed under specific conditions, such
as receiving a person’s ‘informed consent’. Either, an outright ban on
psychometric profile can be considered.

Disinformation

Production of misleading or false content is not limited to the

current historical context, but has always been inherent in human
communication. In the algorithmic public opinion, however, scale,
consequences and responses are arguably more complex and various.
Many recommendations have been provided for this particular
issue which is at the intersection of the systemic and specific issues
hihglighted in this policy review.

Any interventions in the online disinformation space have to

recognise that domestic media and domestic politicians are often part of
some disinformation problems – and, importantly, the public recognises
this and frequently expresses the same level of concern over what they
see as political propaganda and poor journalism as they do over false
and fabricated content (Newman et al. 2018).

Importantly, responses should be fully compliant with the

fundamental principles of freedom of expression, free press and
pluralism, and at the same time future-proof and efficient in averting
public harm. In order to ensure this, the independent High level

The Algorithmic Public Opinion: a Policy Overview 50

 Group on fake news and online disinformation (de Cock Buning,
2018) cautioned against simplistic solutions. More generally, it is
recommended for a multi-dimensional approach that caters for
the need to continually examine the phenomenon and evaluate the
effectiveness of the concrete measures adopted by different actors. At
the same time, it is advocated a self-regulatory approach based on a
clearly defined multi-stakeholder engagement process, framed within
a binding roadmap for implementation, and focused on a set of short
and medium-term actions. There is considerable scope for expanding
and improving the collaborative approach to combating disinformation
by involving all relevant stakeholders (public authorities, platform
companies, private news media, public-service media, and civil society
groups, including factcheckers, media literacy groups, and researchers).

Potential and enacted policies are various and employed at different

levels, actors and part of the process (de Cock Buning, 2018); More
generally, these can be summarized as:

1. Enhance transparency of the digital information ecosystem;

2. Promote and sharpen the use of media and information literacy
approaches to counter disinformation and help users navigate our
digital information environment;
3. Develop tools for empowering users and journalists and foster a
positive engagement with fast-evolving information technologies;
4. Safeguard the diversity and sustainability of the news media eco-
system, and, finally
5. Calibrate the effectiveness of the responses through continuous
research on the impact of disinformation and an engagement
process that includes predefined and time-framed steps combined
with monitoring and reporting requirements.

One of the main concerns concerning disinformation is foreign

influence and propaganda, especially from Russia which has allegedly
used cyber-attacks, disinformation, and financial influence campaigns
to meddle with the internal affairs of most European countries in
order to amplify existing political and social discord and erode trust in
mainstream media and democratic institutions. In particular, elections
have become a target for hybrid-warfare. Although the electoral voting
systems have not been compromised yet, attacks against auxiliary
services have been attempted in some countries. To deal with these
troubling forms of disinformation, there is an urgent need for an official
coordinating body to enable collaboration between private companies
and democratic governments, allowing them to work together to identify
and thwart foreign information operations.

Then, while slow, expensive, and limited in scope, significant

investment in media literacy for citizens of all ages is also likely to
be a key part of increasing societal resilience to various kinds of
disinformation. To make a meaningful difference, media literacy has

The Algorithmic Public Opinion: a Policy Overview 51

 to be a central part of education (as it is already is in some countries)
and significant resources will have to be invested in media literacy for
adults, as a growing body of research suggests that older people may be
both more exposed to disinformation and more likely to share it. For
media and information literacy to be effective, it must be pursued across
teacher training curricula, school curricula, and beyond, and it will
require significant investment and ongoing evaluation and evolution.

Finally, to develop credible and effective policy responses to

disinformation, there is an urgent need for more independent, evidence-
based research. While there is no doubt that there are many and serious
problems of disinformation, we still know little about the scale and
scope in different countries, the actual effects of disinformation, and the
effectiveness of various possible policy interventions.

Conclusions

This policy review has provided an overview of the challenges of

the algorithmic public opinion and the debate on proposed policy
approaches and promising proposals. The algorithmic public opinion
is a complex infrastructure that comprises various algorithms and
these algorithms, in turn, are not discrete tools that inscribe values
and follow specific directives that can be regulated as such, but socio-
technical assemblages that include a variety of actors, stakeholders,
processes and technologies. Ultimately, algorithmic systems represent
only a point of departure to understand the complexity of how public
opinion is disrupted by this infrastructure. To regulate the algorithmic
public opinion means to regulate the environment in which algorithms
perform; this means to understand why they are employed, what are the
ideologies that justify their usage, how algorithms operate in specific
contexts and how they change and are adapted over time, what data
are used for algorithms and how, how users interact with algorithm,
how these are visually represented in interface design, how they are
generally understood, and what are their (unintended) consequences.
Still, the definition of these algorithmic systems is fundamental for an
effective regulation but these are still famously blurred in academia as
well as in policy-making and common usage.

In this policy review, I have analyzed the most significant algorithmic

systems molding the algorithmic public opinion. Arguably, the most
influencing ones are platform recommender systems, the ones that act
as fundamental global gatekeepers allowing content personalization.
The gatekeeping role of personalization algorithms sets up an
unprecedented immense ‘opinion power’ that represents the most
concerning regulatory challenge. Yet, other relevant algorithms have
been discussed; the ones employed in content moderation (to detect
and remove content), social bots (to manipulate content diffusion and
discussions), and, among others, those who help to create, edit and
retrieve content.

The Algorithmic Public Opinion: a Policy Overview 52

 The regulation and governance of an algorithmic public opinion is
certainly unprecedented. This is especially true considering not only
the intrinsic features that make difficult algorithms to research and,
thus, to understand, but also because the data that researchers can
currently access is substantially limited and more often unreliable.
Another important - perhaps obvious - conclusion of this policy review,
in fact, is that it is paramount to allow researchers to access wider
dataset in a legal and transparent manner. This is a condition sine qua non
without which it becomes very difficult to collect convincing evidence
for effective policies as well as to maintain a critical scrutiny over these
powerful algorithms by civil society. The current institutional barriers
to the social data produced in online platforms is indeed a form of
agnotology – the science of ignorance – and the ignorance it produces,
the often ambivalent and constrasting conclusions that scholars draw
from limited datasets, seriously hinders the ability to govern and
regulate this fundamental infrastructure of information societies, while
it has created a perverse dynamic in which researchers could get more
funding and platforms could escape regulation.

The risks that an unregulated algorithmic public opinion pose are

undoubtedly grave, even if scientific evidences are often unconvincing.
These include not only threats to democratic processes and values,
including electoral integrity, but also, more concretely, misinformation,
political polarization, radicalization, and even “addiction”. These can
ultimately lead to collective distraction, confusion and disagreement
that can conduce to dangerous collective inaction in front of epochal
challenges, Covid-19 and climate change. For the most critical, this
could represent an epistemic (and democratic) implosion; in fact, the
pollution of public opinion can lead to public preferences different
to if these were accurately informed, which can have negative policy
implications. The same is true with public opinion more generally,
where policy outputs feed back on public inputs into the policy‐making
process. Therefore, policies for the algorithmic public opinion can even
increase the quality of public policies more generally. This makes their
effectiveness ever more significant.

The question of regulating the algorithmic public opinion is part of

a complex regulatory system involving many different legislative acts
and normative approaches. Policy-making, however, still lacks a shared
vocabulary or frameworks for approaching the incoming challenges
and new models for digital governance will likely need to be developed.
The algorithmic public opinion is particularly challenging to policy-
makers as it intersects with a transformation of media itself, due to
media convergence, media globalization and the rise of global digital
platforms. Legislation in the algorithmic public opinion context always
risks being too broad to be useful, or too specific to be future-proof.

Despite several regulatory attempts in various legal systems, current

regulation and policies are generally still unable to provide adequate
responses to the challenges that the algorithmic public opinion poses.
Rules and policies are fragmented and mostly provide horizontal
approaches to many of these challenges, that is, diverse regulatory areas

The Algorithmic Public Opinion: a Policy Overview 53

 concur to regulate algorithms ultimately reducing their effectiveness.
More generally, the systemic policy challenges analyzed in this policy
review stress that the Internet is a global common infrastructure
rooted in all societies’ economies, politics, and cultures, and that the
tensions it produces are not promptly solvable by the current national
and international legal system. As such, there is the need for a trans-
national governance, especially because the worldwide competing
policy approaches may lead to the fragmentation and, eventually, cyber-
balkanization of the Internet. This realization should lead us to radically
re-think the governance of the Internet itself. At the same time, we
should also avoid naive expectations of the Internet as an ideal public
sphere: there are always new challenges, and human and machines
mistakes will inevitably raise further concerns. It is a never-ending
challenge the governance of the algorithmic public opinion.

The Algorithmic Public Opinion: a Policy Overview 54

 Bibliography

Acquisti, A., Brandimarte, L., & Loewenstein, G. (2015). Privacy and

human behavior in the age of information. Science, 347(6221), 509-514.

Ada Lovelace Institute (2021). Technical methods for regulatory

inspection of algorithmic systems in social media platforms.

Adadi, A., & Berrada, M. (2018). Peeking inside the black-box: A

survey on explainable artificial intelligence (xai). IEEE Access, 6: 52138–
52160, 2018.

Alan Turing Institute, Leslie, D. (2019). Understanding artificial

intelligence ethics and safety: A guide for the responsible design and
implementation of AI systems in the public sector

Albanie, S., Shakespeare, H., & Gunter, T. (2017). Unknowable

manipulators: Social network curator algorithms. arXiv preprint
arXiv:1701.04895.

Algo:aware (2018). Raising awareness on algorithms. Procured by

the European Commission’s Directorate-General for Communications
Networks, Content and Technology.

AlgorithmWatch (2020). Are Algorithms a Threat to Democracy? The

Rise of Intermediaries: A Challenge for Public Discourse. Algorithm
Watch

Amnesty International (2021). The Pegasus Project: https://www.

amnesty.org/en/latest/press-release/2021/07/the-pegasus-project/

Ananny, M., & Crawford, K. (2018). Seeing without knowing:

Limitations of the transparency ideal and its application to algorithmic
accountability. New media & society, 20(3), 973-989.

Ananny, M., & Gillespie, T. (2016). Public platforms: Beyond the cycle
of shocks and exceptions. IPP2016 The Platform Society.

Andreotta, A. J., Kirkham, N., & Rizzi, M. (2021). AI, big data, and the
future of consent. Ai & Society, 1-14.

Backstrom, L. (2013). News feed FYI: A window into news feed (p. 6).
Menlo Park: Facebook for Business.

Balding, C. (2020) ‘Chinese Open Source Data Collection, Big Data,

And Private Enterprise Work For State Intelligence and Security: The
Case of Shenzhen Zhenhua’. Available at SSRN <https://ssrn.com/
abstract=3691999>.

Bank M., Duffy F., Leyendecker V., & Silva M. (2021). The Lobby
Network: Big Tech’s Web of Influence in the EU. Corporate Europe

The Algorithmic Public Opinion: a Policy Overview 55

 Observatory and LobbyControl e.V.

Bernstein, A., de Vreese, C., Helberger, N., Schulz, W., Zweig, K.,
Baden, C., ... & Zueger, T. (2020). Diversity in news recommendations.
arXiv preprint arXiv:2005.09495.

Bietti, E. (2020). From ethics washing to ethics bashing: a view on

tech ethics from within moral philosophy. In Proceedings of the 2020
Conference on Fairness, Accountability, and Transparency (pp. 210-219).

Bogost, I. (2015). The cathedral of computation. The Atlantic, 15(01).

Bodo, B., Helberger, N., Irion, K., Zuiderveen Borgesius, F., Moller,
J., van de Velde, B., ... & de Vreese, C. (2017). Tackling the algorithmic
control crisis-the technical, legal, and ethical challenges of research into
algorithmic agents. Yale JL & Tech., 19, 133.

Borgesius, F. J. Z., Trilling, D., Möller, J., Bodó, B., De Vreese, C. H.,
& Helberger, N. (2016). Should we worry about filter bubbles?. Internet
Policy Review. Journal on Internet Regulation, 5(1).

Borgesius, F. J. Z., Möller, J., Kruikemeier, S., Fathaigh, R. Ó., Irion,

K., Dobber, T., ... & De Vreese, C. (2018). Online political microtargeting:
Promises and threats for democracy. Utrecht Law Review, 14(1), 82-96.

Bösch, C., Erb, B., Kargl, F., Kopp, H., & Pfattheicher, S. (2016). Tales
from the Dark Side: Privacy Dark Strategies and Privacy Dark Patterns.
Proc. Priv. Enhancing Technol., (4), 237-254.

Bozdag, E. (2013). Bias in algorithmic filtering and personalization.

Ethics and information technology, 15(3), 209-227.

Bradford, A. (2020). The Brussels effect: How the European Union rules
the world. Oxford University Press, USA.

Bradshaw, S., & Howard, P. (2017). Troops, trolls and troublemakers:

A global inventory of organized social media manipulation.

Bradshaw, S., Howard, P. N., Kollanyi, B., & Neudert, L. M. (2020).

Sourcing and automation of political news and information over social
media in the United States, 2016-2018. Political Communication, 37(2),
173-193.

Bruns, A. (2019). After the ‘APIcalypse’: social media platforms

and their fight against critical scholarly research. Information,
Communication & Society, 22(11), 1544-1566

Bruns, A. (2019). It’s not the technology, stupid: How the ‘Echo
Chamber’and ‘Filter Bubble’metaphors have failed us. International
Association for Media and Communication Research.

Bunting, M. (2018). From editorial obligation to procedural

The Algorithmic Public Opinion: a Policy Overview 56

 accountability: policy approaches to online content in the era of
information intermediaries. Journal of Cyber Policy, 3(2), 165-186.

Bucher, T. (2017). The algorithmic imaginary: Exploring the ordinary

affects of Facebook algorithms. Information, Communication & Society,
20(1), 30–44.

Burri, M. (2015). Contemplating a’Public Service Navigator’: In Search

of New (and Better) Functioning Public Service Media. International
Journal of Communication, 9, 1341-1359.

Byung-Chul, H. (2016). Psicopolitica. Il neoliberismo e le nuove tecniche

del potere. Nottetempo, Milano.

Cabral, L. (2021). Merger policy in digital industries. Information

Economics and Policy, 54, 100866.

Carr, N. (2010). The shallows: How the internet is changing the way we
think, read and remember. Atlantic Books Ltd.

Castro, J. C. L. (2016). Social networks as dispositives of neoliberal

governmentality. Journal of Media Critiques, 2(7), 85–102.

Cobbe, J., & Singh, J. (2019). Regulating Recommending: Motivations,

Considerations, and Principles. Considerations, and Principles (April 15,
2019).

Cobbe, J. (2020). Algorithmic censorship by social platforms: power and

resistance. Philosophy & Technology, 1-28.

Cwajg C. M. (2020). Transparency Rules in Online Political

Advertising: Mapping Global Law and Policy.

Dalton, R. J., & Klingemann, H. D. (2007). Citizens and political

behavior. In The Oxford handbook of political behavior.

DECODE, Vercellone, C., Brancaccio, F., Giuliani, A., Puletti, F.,

Rocchi, G., & Vattimo, P. (2018). Data-driven disruptive commons-based
models (Doctoral dissertation, CNRS).

Deibert, R.J. (2019). The road to digital unfreedom: three painful

truths about social media. J. Democracy, 30(1), 25–39.

DeNardis, L., & Hackl, A. M. (2015). Internet governance by social

media platforms. Telecommunications Policy, 39(9), 761-770.

DeVito, M. A. (2017). From editors to algorithms: A values-based

approach to understanding story selection in the Facebook news feed.
Digital Journalism, 5(6), 753–773.

Diakopoulos, N. (2014). Algorithmic Accountability. Digital

Journalism, 3(3), 398–415. doi:10.1080/21670811.2014.9764

The Algorithmic Public Opinion: a Policy Overview 57

 Diakopoulos, N. (2016). Accountability in algorithmic decision
making. Communications of the ACM, 59(2), 56-62.

Dobber, T., Ó Fathaigh, R., & Zuiderveen Borgesius, F. (2019). The

regulation of online political micro-targeting in Europe. Internet Policy
Review, 8(4).

EDRi (2021). Beyond Debiasing Regulating AI and its inequalities.

EDRi: https://edri.org/wp-content/uploads/2021/09/EDRi_Beyond-Debiasing-
Report_Online.pdf

Edwards, L., & Veale, M. (2017). Slave to the algorithm: Why a right to
an explanation is probably not the remedy you are looking for. Duke L. &
Tech. Rev., 16, 18

Epstein, R., & Robertson, R. E. (2015). The search engine

manipulation effect (SEME) and its possible impact on the outcomes
of elections. Proceedings of the National Academy of Sciences, 112(33),
E4512-E4521.

de Cock Buning, M. (2018). A multi-dimensional approach to

disinformation: Report of the independent High level Group on fake
news and online disinformation. Publications Office of the European
Union.

Fallis, D. (2020). The Epistemic Threat of Deepfakes. Philosophy &

Technology, 1-21.

Finck, M. (2018). Digital co-regulation: designing a supranational

legal framework for the platform economy. European Law Review

Floridi, L. (2016). Tolerant paternalism: Pro-ethical design as a

resolution of the dilemma of toleration. Science and Engineering Ethics,
22(6), 1669–1688.

Floridi, L. (2018). Soft ethics and the governance of the digital.

Philosophy & Technology, 31(1), 1-8.

Floridi, L. (2021a). Translating principles into practices of digital

ethics: Five risks of being unethical. In Ethics, Governance, and Policies in
Artificial Intelligence (pp. 81-90). Springer, Cham.

Floridi, L. (2021b). The End of an Era: from Self-Regulation to Hard

Law for the Digital Industry. Philosophy & Technology, 1-4.

Fjeld, J., Achten, N., Hilligoss, H., Nagy, A., & Srikumar, M. (2020).
Principled artificial intelligence: Mapping consensus in ethical and
rights-based approaches to principles for AI. Berkman Klein Center
Research Publication, (2020-1).

Fogg, B. J., Lee, E., & Marshall, J. (2002). Interactive technology and
persuasion. The Handbook of Persuasion: Theory and Practice. Thousand

The Algorithmic Public Opinion: a Policy Overview 58

 Oaks: Sage

Fuchs, C., & Sandoval, M. (2015). The political economy of capitalist

and alternative social media. In The Routledge companion to alternative
and community media (pp. 165-175). Routledge

Gal, M. S. (2017). Algorithmic challenges to autonomous choice.

Michigan Telecommunications and Technology Law Review, 2017.

Gillespie, T. (2014). The relevance of algorithms. Media technologies:

Essays on communication, materiality, and society, 167(2014), 167.

Gillespie, T. (2018), ‘Regulation of and by platforms’, in J. Burgess, A.

Marwick and T. Poell (eds), The SAGE Handbook of Social Media, London:
SAGE, pp. 254–78.

Gillespie, T., Aufderheide, P., Carmi, E., Gerrard, Y., Gorwa, R.,
Matamoros-Fernández, A., ... & West, S. M. (2020). Expanding the debate
about content moderation: Scholarly research agendas for the coming
policy debates. Internet Policy Review, 9(4), 4 1, 29.

Google, ‘Effect of disabling third-party cookies on publisher revenue’

(2020). https://services.google.com/fh/files/misc/disabling_third-party_
cookies_publisher_revenue.pdf> accessed 13 November 2020.

González-Cabañas, J., Cuevas, Á., Cuevas, R., López-Fernández, J.,

& García, D. (2021, November). Unique on Facebook: formulation and
evidence of (nano) targeting individual users with non-PII data. In
Proceedings of the 21st ACM Internet Measurement Conference (pp. 464-479).

Gorwa, R. (2019). What is platform governance?. Information,

Communication & Society, 22(6), 854-871.

Gorwa, R., Binns, R., & Katzenbach, C. (2020). Algorithmic content

moderation: Technical and political challenges in the automation of
platform governance. Big Data & Society, 7(1), 2053951719897945.

Gorwa, R., & Guilbeault, D. (2020). Unpacking the social media bot: A
typology to guide research and policy. Policy & Internet, 12(2), 225-248.

Gray, C. M., Kou, Y., Battles, B., Hoggatt, J., & Toombs, A. L. (2018).
The dark (patterns) side of UX design. In Proceedings of the 2018 CHI
Conference on Human Factors in Computing Systems (p. 534). ACM.

Helberger, N., Pierson, J., Poell, T. (2018). Governing online

platforms: from contested to cooperative responsibility. Inf. Soc. 34(1),
1–14.

Helberger, N., Eskens, S. J., van Drunen, M. Z., Bastian, M. B., &
Möller, J. E. (2019). Implications of AI-driven tools in the media for
freedom of expression.

The Algorithmic Public Opinion: a Policy Overview 59

 Helberger, N. (2019). On the democratic role of news recommenders.
Digital Journalism, 7(8), 993-1012.

Helberger, N. (2020). The political power of platforms: How current

attempts to regulate misinformation amplify opinion power. Digital
Journalism, 8(6), 842-854.

Hildén, J. (2021). The Public Service Approach to Recommender

Systems: Filtering to Cultivate. Television & New Media,
15274764211020106.

HLEG AI (High-Level Expert Group on Artificial Intelligence). (2019).

Ethics guidelines for trustworthy AI.

Hildebrandt, M., & Koops, B. J. (2010). The challenges of ambient law

and legal protection in the profiling era. The Modern Law Review, 73(3),
428-460.

Hildebrandt, M. (2015). Smart technologies and the end (s) of law: novel
entanglements of law and technology. Edward Elgar Publishing.

Hildebrandt, M. (2022). The Issue of Proxies and Choice

Architectures. Why EU law matters for recommender systems. Frontiers
in Artificial Intelligence, 73.

Hirsch, D. D. (2010). The law and policy of online privacy: Regulation,

self-regulation, or co-regulation. Seattle UL Rev., 34, 439.

Hoffmann, C. P., Lutz, C., Meckel, M., & Ranzini, G. (2015). Diversity
by choice: Applying a social cognitive perspective to the role of public
service media in the digital age. International Journal of Communication,
9(1), 1360–1381.

Hwang, T., (2020) Subprime Attention Crisis: Advertising and the Time
Bomb at the Heart of the Internet. Farrar Straus & Giroux. Rome

De La Chapelle, B. and L. Porciuncula (2021). We Need to Talk

About Data: Framing the Debate Around Free Flow of Data and Data
Sovereignty. Internet and Jurisdiction Policy Network.

Kahneman, D. (2011). Thinking, fast and slow. Macmillan.

Kay, M., Matuszek, C., & Munson, S. A. (2015). Unequal

representation and gender stereotypes in image search results for
occupations. In Proceedings of the 33rd Annual ACM Conference on Human
Factors in Computing Systems (pp. 3819-3828)

Kaye, D. 2019. Speech Police: The Global Struggle to Govern the Internet.
New York: Columbia Global Reports

Keymolen, E. (2016). Trust on the line: a philosophycal exploration of

trust in the networked era

The Algorithmic Public Opinion: a Policy Overview 60

 Kidron, B., Evans, A., Afia, J., Adler, J. R., Bowden-Jones, H., Hackett,
L., & Scot, Y. (2018). Disrupted childhood: The cost of persuasive design.
5Rights.

Kingaby H., and Kaltheuner F. (2020). ‘Ad Break For Europe The
Race To Regulate Digital Advertising And Fix Online Spaces’. Mozilla
Foundation.

Kramer, A. D., Guillory, J. E., & Hancock, J. T. (2014). Experimental

evidence of massive-scale emotional contagion through social networks.
Proceedings of the National Academy of Sciences, 111(24), 8788-8790.

Leerssen, P. (2020). The Soap Box as a Black Box: Regulating

transparency in social media recommender systems. European Journal of
Law and Technology, 11(2).

Leerssen, Paddy: Platform research access in Article 31 of the Digital

Services Act: Sword without a shield?, VerfBlog, 2021/9/07, https://
verfassungsblog.de/power-dsa-dma-14/, DOI: 10.17176/20210907-214355-
0.

Lessig, L. (1999). Code: And other laws of cyberspace. Basic Books.

Loecherbach, F., Moeller, J., Trilling, D., & van Atteveldt, W. (2020).
The unified framework of media diversity: A systematic literature
review. Digital Journalism, 8(5), 605-642.

Luguri, and Strahilevitz (2021). Committee for the Study of Digital

Platforms Privacy and Data Protection Subcommittee. Stiegler Report.

Masood, M., Nawaz, M., Malik, K. M., Javed, A., & Irtaza, A. (2021).
Deepfakes Generation and Detection: State-of-the-art, open challenges,
countermeasures, and way forward. arXiv preprint arXiv:2103.00484.

Milano, S., Taddeo, M., & Floridi, L. (2020). Recommender systems

and their ethical challenges. Ai & Society, 35(4), 957-967.

Moen, G. M., Ravna, A. K., & Myrstad, F. (2018). Deceived by design.

Forbrukerrådet.

Morley, J., Floridi, L., Kinsey, L., & Elhalal, A. (2019). From What to
How: An Initial Review of Publicly Available AI Ethics Tools, Methods
and Research to Translate Principles into Practices. Science and
Engineering Ethics, 1-28.

Morozov, E. (2017). “Moral panic over fake news hides the real enemy
– the digital giants”, The Guardian, Jan 08. https://www.theguardian.
com/commentisfree/2017/jan/08/blaming-fake-news-not-the-
answerdemocracy-crisis

The Algorithmic Public Opinion: a Policy Overview 61

 Mosseri, A. 2018. News feed FYI: Bringing people closer together.
https://www.facebook.com/business/news/news-feed-fyi-bringingpeople-closer-
together

Nahmias, Y., & Perel, M. (2021). The Oversight of Content Moderation

by AI: Impact Assessments and Their Limitations. Harv. J. on Legis., 58,
145.

Napoli, P. M. (1999). Deconstructing the diversity principle. Journal of

communication, 49(4), 7-34.

Napoli, P., & Caplan, R. (2017). Why media companies insist they’re
not media companies, why they’re wrong, and why it matters. First
Monday.

Napoli, P. M. (2019). User data as public resource: Implications for

social media regulation. Policy & Internet, 11(4), 439-459.

Nash, V. (2019). Internet Regulation and the Online Harms White

Paper: Stakeholder Workshop Summary. Available at SSRN 3412790.

Neumann N., Tucker C. E., and Whitfield T., ‘Frontiers: How effective
is third-party consumer profiling? Evidence from field studies’ (2019)
38(6) Marketing Science, 918.

Newman, N., Fletcher, R., Kalogeropoulos, A., Levy, D. A., & Nielsen,
R. K. (2016). Digital news report 2016. Reuters Institute for the Study of
Journalism.

Newman, N. (2020). Journalism, media, and technology trends and

predictions. Reuters Institute for the Study of Journalism.

Pariser, E. (2011). The Filter Bubble: How the New Personalized Web is
Changing What We Read and How We Think. Penguin, Westminster.

Pasquale, F. (2015). The black box society: The secret algorithms that
control money and information. Harvard: Harvard University Press.

Rahman, K. S. (2018). Regulating informational infrastructure:

Internet platforms as the new public utilities. Georgetown Law and
Technology Review, 2, 2.

Reviglio, U. (2019). Serendipity as an emerging design principle of

the infosphere: challenges and opportunities. Ethics and Information
Technology, 21(2), 151-166.

Reviglio, U., & Agosti, C. (2020). Thinking outside the Black-box: The
case for “algorithmic sovereignty” in social media. Social Media+ Society,
6(2), 2056305120915613.

Reviglio, U. (2022). The untamed and discreet role of data brokers in

surveillance capitalism: a transnational and interdisciplinary overview.

The Algorithmic Public Opinion: a Policy Overview 62

 Internet Policy Review, 11(3).

Ricci, F., Rokach, L., & Shapira, B. (2015). Recommender systems:

introduction and challenges. In Recommender systems handbook (pp. 1-34).
Springer, Boston, MA.

Romer, P. (2021, May 17). Taxing Digital Advertising. adtax.paulromer.net.

Scasserra, S., & Elebi, C. M. (2021). Digital colonialism Analysis of

Europe’s trade agenda. Policy.

Scott Morton, F., Bouvier, P., Ezrachi, A., Jullien, B., Katz, R.,
Kimmelman, G., ... & Morgenstern, J. (2019). Committee for the study of
digital platforms: Market structure and antitrust subcommittee report.
Stigler Center for the Study of the Economy and the State, University of
Chicago Booth School of Business.

Shenkman, C., Thakur, D., & Llansó, E. (2021). Do You See What I
See?. Executive Summary. Policycommons.net.

Statista Research Department, Nov 1, 2021: https://www.statista.com/

statistics/264810/number-of-monthly-active-facebook-users-worldwide/

Sunstein, C.R. (2017). #Republic: Divided Democracy in the Age of Social

Media. Princeton University Press, Princeton. 4.

Taddeo, M. (2012). Information warfare: A philosophical perspective.

Philosophy & Technology, 25(1), 105-120.

Taylor, L., Floridi, L., & Van der Sloot, B. (Eds.). (2016). Group privacy:
New challenges of data technologies (Vol. 126). Springer.

Thaler, R. H., & Sunstein, C. R. (2019). Nudge: Improving Decisions

about Health, Wealth, and Happiness. Penguin.

Thurman, N. (2011). Making ‘The Daily Me’: Technology, economics

and habit in the mainstream assimilation of personalized news.
Journalism, 12(4), 395–415

Thurman, N., Moeller, J., Helberger, N., & Trilling, D. (2019). My

friends, editors, algorithms, and I: Examining audience attitudes to news
selection. Digital Journalism, 7(4), 447-469.

Tucker, J. A., Guess, A., Barberá, P., Vaccari, C., Siegel, A., Sanovich,
S., ... & Nyhan, B. (2018). Social media, political polarization, and
political disinformation: A review of the scientific literature. Political
polarization, and political disinformation: a review of the scientific literature.

Vaccari, C., & Chadwick, A. (2020). Deepfakes and disinformation:

Exploring the impact of synthetic political video on deception,
uncertainty, and trust in news. Social Media+ Society, 6(1).

The Algorithmic Public Opinion: a Policy Overview 63

 Verduyn, P., Ybarra, O., Résibois, M., Jonides, J., & Kross, E. (2017).
Do social network sites enhance or undermine subjective well-being? A
critical review. Social Issues and Policy Review, 11(1), 274–302.

Wardle, C., & Derakhshan, H. (2017). Information disorder: Toward

an interdisciplinary framework for research and policy making. Council
of Europe report, 27, 1-107.

Wieringa, M. (2020, January). What to account for when accounting

for algorithms: a systematic literature review on algorithmic
accountability. In Proceedings of the 2020 conference on fairness,
accountability, and transparency (pp. 1-18).

World Bank (2021). World Development Report 2021: Data for Better
Lives. Washington, DC: World Bank.

Yesilada, M., & Lewandowsky, S. (2021). A systematic review: The

YouTube recommender system and pathways to problematic content.

Yeung, K. (2017). ‘Hypernudge’: big data as a mode of regulation by

design. Inf. Commun. Soc. 20 (1), 118–136.

Youyu, W., Kosinski, M., & Stillwell, D. (2015). Computer-based

personality judgments are more accurate than those made by humans in
Proceedings of the National Academy of Sciences, 112(4), 1036-1040.

Zarsky, T.Z. (2019). Privacy and manipulation in the digital age. Theor. Inq.
Law 20(1), 157–188.

Zuboff, S. (2015). Big other: surveillance capitalism and the prospects of an

information civilization. Journal of information technology, 30(1), 75-89.

The Algorithmic Public Opinion: a Policy Overview 64

 Box 1 — The Reductionism of Profiling Technologies

Profiling technologies that allow personalization and recommendations create a kind of

knowledge that is inherently probabilistic. As a matter of fact, profiling technologies cannot
produce or detect a sense of self. Data used are derived only from what is actually observed and
are a small subset of possibly observed behaviors. Then, this narrow subset of recorded behaviors
must be converted into a digital representation. In this process, information may be lost or
misinterpreted. Eventually, four key characteristics of profiling technologies arise:

→ Profiling is highly behaviorist its assumptions. Behaviorism seeks to eliminate theoretical

causal mechanisms of human behavior (beliefs, intentions, goals, etc.) and focus instead on
what can be observed measured and recorded. Behaviorism actually “confuses the modest
proposition that we only have access to what we can observe with the claim that only that
which we can observe exists and/or matters” (Hildebrandt, 2021, p.7). As such, behaviors
most amenable to measurement tend to be recorded whereas what is inside the individual
is mostly ignored. Profiling focuses on predicting a very narrow set of possible behaviors,
often limited by the context of the application. This can make it seem more powerful and
accurate than it really is, especially when predictive performance is evaluated. Most of these
systems, for example, tend to capture mostly positive feedback.
→ Profiling is meant to infer users’ “preferences”, roughly meaning “what people want.” This,
however, has been justified by the assumption that people always choose what they want,
an idea from 20th-century economics called “revealed preferences”. Not only this approach
is disputble but it can also lead to a variety of unwanted outcomes including clickbait,
addiction, or algorithmic manipulation.
→ Profiling becomes ever more persuasive thanks to the emergence of psychographic
techniques and “emotional AI”. These techniques, however, often rely on a highly contested
scientific paradigm that argues that all humans, everywhere, experience the same basic
emotions, and express those emotions in the same way. Those emotions include happiness,
anger, sadness, disgust, surprise, and fear. This paradigm of universal emotions is
insufficiently evidence-based and poorly regarded in the relevant scientific communities.
→ Profiling uses data not only from the individual user, but also from other users. This is clear
in the case of RSs using social network data (e.g. collaborative filtering). This lead to question
the extent to which recommendations are actually personalized for individuals.

The underlined – somehow inevitable – reductionism is concerning for it can undermine the
development of the self, the formation of a healthy and informed public opinion and ultimately
serve only the interests of service providers. These limitations need to be acknowledged and
eventually overcome.

The Algorithmic Public Opinion: a Policy Overview 65

 Box 2 — The Case of Facebook’s Algorithm

Facebook has 2.85 billion active users (Statista, 2021) and is likely the most pervasive and powerful
intermediary on the Internet. Its unique relevance depends also on the fact that regarding
personalization, search engines like Google are, to some extent, less problematic because they
tend to deliver more one-size-fits-all services. The system of news updating in Facebook – the
NewsFeed – has an ever growing central role in the global information flow. On average, it selects
200 posts on 2.000 recommendable posts each day, therefore hiding 90% of the content. It is based
on EdgeRank, a complex algorithm which constantly changes its outputs. In a rather unique
analysis, DeVito (2017) examined the few publications by Facebook that provided information
about the operating principles of the algorithm (e.g., press releases, blogs or patents) and was
able to identify nine relevant “editorial values”: “Friend relationships, explicitly expressed user
interests, prior user engagement, implicitly expressed user preferences, post age, platform
priorities, page relationships, negatively expressed preferences, and content quality” (DeVito,
2017, p. 14). In this context, of particular interest is their relative importance: the ranking criteria
of Facebook show only slight overlaps with classical news factors: only the prioritization of new
stories and local stories are related to the news factors “novelty” and “proximity” (Napoli, 2019).
However, the three most important characteristics mentioned are the quality of the relationships
(affinity) as well as explicitly and implicitly stated user interests, the latter being identified by the
user’s previous behavior. The analysis of Facebook’s patents in particular showed that, “friend
relationships are a guiding value that mediates the application of all the other values” (DeVito,
2017, p. 14). This has been emphasized in an update of the news feed in 2018 when Facebook
explicitly strengthened the prominence of posts assumed to stimulate discussions and other
“meaningful interactions” such as shares and likes in user’s networks (Mosseri, 2018). Ultimately,
the information available on the selection principles of the Facebook algorithm is only an
approximation of the actual selection, as the more than 200 selection criteria are well-guarded
company secrets and constantly updated and adapted (De Vito, 2017).

The Algorithmic Public Opinion: a Policy Overview 66

 Box 3 — The Challenges of Media Diversity

As a reaction to the new mostly personalized media environment, media scholars and policy-
makers discussed how to preserve and cultivate media diversity. Generally, this is achieved when
users autonomously enjoy a diverse media diet. Yet, diversity is not a simple clear-cut concept.
Napoli (1999) provided a useful taxonomy, dividing diversity into three categories: source,
content, and exposure diversity. Source diversity refers to both the plurality of media sources,
their ownership as well as the diversity of the workforce at the media organizations. Content
diversity refers to the diversity of programing, the diversity of represented demographic groups
and diversity of opinion. Lastly, exposure diversity refers to the extent citizens consume diverse
programing. This gained prominence online as the focus of policymakers and academics has
partly shifted from spheres of production to those of distribution. And even if the exposure
to diversity online is generally more than in traditional media, it does not always end up in
an ‘experience of diversity’ (Hoffmann et al., 2015). Cognitive and affective factors that drive
Internet users must also be considered. This requires to employ a user-centric perspective and
extend beyond the assumption that supply and exposure diversity equals experience of diversity,
and that diversity of sources equals diversity of content. As such, scholars like Helberger et al.
(2016) suggest that PSM should employ RSs that expose their users to diverse content. But, still,
how much exposure to how many different contents and sources can be considered sufficient?
Fundamentally, the problem lies on the conceptualization of diversity. Media diversity, in
particular, is a rich and complex ideal that can be achieved in many different ways, and its
interpretation differs significantly per discipline (Loecherbach et al., 2020; Bernstein et al., 2020).

While many scholars from different disciplines agree that media diversity is an important value
that we should include even in the design of institutions, policies and online services, this
value is often reduced to single definitions such as “source diversity” or “hearing the opinion of
the other side”. There is a need for a more detailed normative conceptualization of this value.
Only then, perhaps, it will be possible to translate this complex value into design requirements
of information intermediaries and move towards viable solutions that can be implemented.
Importantly, there are tensions (trade-offs) in diverse design proposals, for example ‘diversity’ and
‘trustworthiness’ on the one hand, and ‘non-discrimination’ and ‘neutrality’ on the other. While
the former requires algorithmic systems to prioritize certain content, the latter could arguably
prohibit the drawing of such distinctions (e.g. conspiracy theories increase diversity). Also,
democracy theories have conflicting values (Helberger, 2019). Thus, the issue of diversity and
its design cannot be ‘solved’ objectively or definitively, rather, throughout more interdisciplinary
experimentations and mutidisciplinary collaborations it will eventually be possible to weight
diversity in design and recommendation algorithms and ultimately approximate a more diverse
media consumption.

The Algorithmic Public Opinion: a Policy Overview 67

 Box 4 — The Rise of Deepfakes

Deepfakes (a portmanteau of “deep learning” and “fake”) are synthetic media in which a person
in an existing image, video or even audio is replaced with someone else’s likeness. In the last few
years, they have captured widespread attention for their uses in celebrity pornographic videos,
revenge porn and fake news. Due to increased interests from amateurs, commercial companies
and even governments, along with video generation easiness and improved video quality, they
elicited responses from both industry and government to detect and limit their use as well as from
scholars to understand their potential impact (Fallis, 2020; Vaccari and Chadwick, 2020; Masood
et al, 2021). Still, research is limited and policies are lacking.

Deepfakes mainly threaten to worsen epistemic uncertainty and, as a consequence, to decrese

the quality of public opinion’s formation. So far, however, most deepfakes currently present on
social media platforms may be regarded as harmless, entertaining, or artistic. There are, however,
also some examples where deepfakes have been used for revenge porn, hoaxes, for political or
non-political influence, and even financial fraud. They undoubtedly have the ability to harm,
increase misinformation and epistemic uncertainty while at the same time decrease trust over
news, people, politicians, and among foreign governments. Deepfakes have indeed the potential to
initiate political tension, conflicts, violence, and even war worldwide. In the worst case scenario,
deepfakes may trigger or substantially influence an informational cold warfare for they could
sway political elections and initiate or worsen geopolitical conflicts, making people ever more
cynical and distrustful and politicians eager to restore “order” and “certainty” through illiberal
policies curtailing free speech and other civil rights.

Raising awareness among the population and systems of prevention, detection and control could
eventually result sufficient to tame the threats of deepfakes (Masood et al, 2021). As Vaccari and
Chadwick (2020) argue, it is also possible that the reduction of trust in news on social media
resulting from the uncertainty induced by deepfakes may not generate cynicism and alienation,
but skepticism. Skepticism can reduce susceptibility to misinformation effects if it prompts people
to question the origins of information that may later turn out to be false while at the same time
ensuring that accurate information is recognized and valued. While skepticism is no panacea, it is
much less problematic for democracy than cynicism and may be a sign, or even a component, of a
healthily critical but engaged online civic culture.

The Algorithmic Public Opinion: a Policy Overview 68

 Box 5 — The Role of European Union

As the United States remains paralysed by partisanship and the concern to maintain its hegemony,
and as the Chinese government aggressively pursues its own distinct approach to digital media
based on very different values, the European Union (EU) is increasingly emerging as a global
policy entrepreneur on digital issues. As the UN Special Rapporteur on freedom of opinion and
expression David Kaye has repeatedly pointed out, from the more active informal and formal
regulation of online content to more robust competition and data protection policies, Europe will
de facto regulate the global internet (Kaye, 2019), in line with the ‘brussels effect’ which refers to
the European global influence on markets setting standards by law, in particular due to the strong
Europeans’ citizens purchase power (Bradfort, 2020).

As policymakers all over the world look to Europe for inspiration, this is a unique opportunity for
the EU and its member states to show leadership and demonstrate what truly democratic digital
media policies can look like. The EU has indeed set its overarching ambition on a human-centric
approach to AI. This endeavior has been captured in the notion of Trustworthy AI, which the
High Level Expert Group on Artificial Intelligence characterised in terms of three components
– being lawful, ethical and robust – and in line with the core tenets of the European Union:
fundamental rights, democracy and the rule of law. The Ethics Guidelines for Trustworthy AI
(2019a) also constitutes a crucial first step in delineating the type of AI that EU wants and do not
want. Another related and relevant policy report (2019b) presents a set of policy and investment
recommendations on how AI can actually be developed, deployed, fostered and scaled in the EU,
all the while maximising its benefits whilst minimising and preventing its risks. For this purpose,
they formulated a number of concrete recommendations addressed to the European Institutions
and Member States.

There are, of course, several challenges for the leading role of EU in the digital legal arena; firstly,
to survive against ever more powerful American and Chinese platforms, EU needs to foster
European, cross border platforms and allow European media innovations to the scale and achieve
the potential of the Digital Single Market (Klossa, 2019). Secondly, the policy process is rather
slow; there is a high risk that the innovative regulatory impetus we see in recent draft proposals
will be outdated before they are even implemented. Thirdly, there are strong lobbying activities
from big tech that require more transparency regulation and public oversight (Bank et al., 2021).
Fourthly, even the EU can be criticized to be another actor in what is called “digital colonialism”,
particularly in the African continent (Scasserra and Elebi, 2021). All in all, EU has the potential to
represent a third innovative pole between US and China but undoubtedly faces serious limits and
challenges ahead.

The Algorithmic Public Opinion: a Policy Overview 69

 Box 6 — Public Service Media and Personalization Systems

Public Service Media (PSM) is the digital version of Public Service Broadcasting (PSB) and it is
defined as a universally available media service. These institutions are especially developed in
Europe compared, for example, to the US. Nowadays a majority of PSM in Europe are currently
moving in the direction of digital and algorithmic personalization (Van den Bulck and Moe, 2017),
assuming the role of a “Public Service Navigator”, meaning a mechanism for influencing the
conditions of access to content, particularly its visibility, discoverability, and usability

(Burri, 2015).

Actual specifications for RSs, however, are still to be developed. Due to fears of filter bubbles and
echo chambers, the attention of the EBU has been mostly focused on exposure to diversity — at
the cost of considering the need for common arenas of discourse, thus the values of universality
and publicness, and how broadcasting and interactivity/choice can reinforce each other. There
are in fact a number of trade-offs involved in personalization systems that European PSM have to
confront with. Many institutions with similar histories and comparable media system frameworks
are taking up different positions. Some consider the possibility to reach the value of universality
through personalization, while other consider personalization to work against it. Some privilege
implicit personalization over explicit personalization, with very different outcomes on PSM goals.
Another problem is that PSM cannot readily rely on ready-made RSs, due to the fact that most RSs
are commercial and generally promote consumption over other values. They must build systems
of their own or modify existing systems. And given that the former is very challenging, it is
becoming more common that RSs are provided by external contractors (Hildén, 2021).

At the same time, it becomes ever more difficult to compete with commercial RSs which
are engagement-driven. The actual risk is to lose ever more audience. Nonetheless, PSM is
generally taking the challenge, which is also an opportunity. PSM could provide affordances and
tools for users – especially in social media – according to its traditional, democratic values, yet
with updated goals and strategies. PSM could thus help to set the standards on a more mature,
democratic and not profit-driven information and news consumption. Eventually, it could
contribute to solve the main challenges to experience diversity online (see Hoffmann et al., 2015).
Yet, this seems unlikely as PSM is composed of several actors that rarely act in concert while
mainstream social media would have all the sufficient resources, know-how and reach to develop
more effectively RSs driven by PSM values.

The Algorithmic Public Opinion: a Policy Overview 70

 Box 7 — The Role of Design

The role of algorithms in affecting public opinion cannot be understood without considering
how users interact with the same algorithms, and this usually occurs through design. In fact,
not only design choices allow users to explicitly influence their online experience, for example
for personalization, but, more broadly, these also influence the online debate as well as the
dissemination, production and consumption of information. Obvious examples are the ‘like’ or the
‘share’ button. These features are clearly not just features but they also carry symbols and steer
behaviors. Design choices and the affordances they allow are paramount to the functioning of the
algorithmic public opinion. By clicking and liking users fuel the algorithms, which in their turn
generate the information flows fed back to users. Moreover, the clicks and likes fuel the interest
and engagement of developers, researchers and advertisers who help to keep the platforms in
business. In this context, an ever more influential role is assumed by ‘deceptive design’ or ‘dark
patterns’. “Dark patterns” define instances where designers use their knowledge of human
behavior (e.g., psychology) and the desires of end-users to implement deceptive functionality
that is not in the user’s best interest (Gray et al., 2018)20. U.S. Federal Trade Commissioner Rohit
Chopra recently defined dark patterns as “design features used to deceive, steer, or manipulate
users into behavior that is profitable for an online service, but often harmful to users or contrary
to their intent.” These are employed extensively not only during terms of conditions and privacy
updates (Moen et al., 2018), but also for what concerns personalization and its features. A helpful
taxonomy, developed by Christoph Bösch and co-authors (2016), identifies classic types of privacy
dark patterns including bad defaults (which we propose a framework for identifying above, and
one example of which is a choice between “Yes” and “Not Now” rather than “Yes” and “No”),
privacy zuckering (i.e., providing users options to adjust their privacy settings that are needlessly
complex, granular, or confusing), forced account registration (seeming to require registration to
use a service), hidden fees or terms added at the end of a long transaction (how did that wind up in
my online shopping cart?), forced account preservation (making it impossible to delete accounts
once created), and address book leeching (requesting users’ contacts at the time of activation
and then spamming users’ contacts with email invitations) . Design can also help to nudge users
without introducing manipulative measures. Such approach is embraced by the philosopher
of information Luciano Floridi (2016) who advocates for what he calls ‘pro-ethical design’ or
“tolerant paternalism” that, in short, it is the attempt to modify the level of abstraction of the
choice architecture by educating users to make their own critical choices and to assume explicit
responsibilities. This approach is mainly grounded on the nudge theory (Thaler and Sunstein,
2019) and ultimately based on behavioralist assumptions (see Box 1 — The Reductionism of Profiling
Technologies) which are debatable whether they are right and even democratic at all (Hildebrandt,
2021).

To conclude, in the regulation and governance of algorithms the role of design cannot be
ignored. Policy-makers and designers are increasingly aware of this; “design policy” is a new area
that looks at the role of design in products and software and then analyzes how design operate
in relation to policy and technology. In other words, it is the act of using design to make policies
around software and hardware understandable. Design policy indeed recognizes that design
affects technology, including how technology looks and feels to consumers, and what consumers
can ‘see’ or know about technology.

The Algorithmic Public Opinion: a Policy Overview 71

 Box 8 — The Role of Ethics

In the struggle of policy-makers to catch up with the rapid pace of technological innovation,
ethics acquires an ever more essential role of moral evalutation so as to complement and improve
digital and AI governance. The increasing demand for reflection and clear policies on the impact
of AI on society has produced several initiatives that state principles, values, or tenets to guide
the development and adoption of AI, including algorithms. Their promise is to condense complex
ethical considerations or requirements into formats accessible to a significant portion of society,
including both the developers and users. The inventory of AI ethics guidelines compiled by
Algorithm Watch lists over 160, of which 88% having been released after 2016, and is constantly
being updated. An obvious risk is unnecessary repetition and overlap, if the various sets of
principles are similar, or confusion and ambiguity, if they differ.

Analyzing 36 sets of AI principles Fjeld et al. (2020) noted a ‘convergence’ around 8 key themes
shared by the various sets of principles: Privacy, Accountability, Safety and Security, Transparency
and Explainability, Fairness and Non-discrimination, Human Control of Technology, Professional
Responsibility and Promotion of Human Values. While we could see this apparent convergence as
a sign that we are moving towards some common ethical foundation, we need to acknowledge that
it is far easier for companies and governments to sign up to relatively vague ethical principles than
it is for them to change business practices or enforce restrictive laws. If principles committing
companies to fairness and non-discrimination had serious legal consequences, there should have
been far more heated disagreement about precisely how to define these principles. Concepts such
as fairness in machine learning are not only contentious - there are indeed 21 different definitions!
- but researchers have proven that different definitions of fairness are in fact mutually exclusive.

There are two major risks arising in this context; on the one hand, ‘ethics bluewashing’, as the
practice of fabricating or exaggerating a company’s interest in equitable AI systems that work for
everyone” (Floridi, 2021a, p.3). Ethics guidelines thus would serve as means to dodge regulation.
On the other hand, there is the risk of “ethics bashing” as “a tendency, common amongst social
scientists and non-philosophers, to trivialize “ethics” and “moral philosophy” by reducing more
capacious forms of moral inquiry to the narrow conventional heuristics or misused corporate
language they seek to criticize” (Bietti, 2020, p.221). Grappling with the role of philosophy and
ethics in tech policy requires moving beyond both ethics washing and ethics bashing and seeing
ethics as a mode of inquiry.

Clear shifts in governmental policy which can be directly traced back to preceding and
corresponding sets of AI Ethics Principles, however remain few and far between. Yet, as
governmental policy-making takes time, it may simply be premature to gauge (or dismiss)
their impact. It is true, however, that most attention for all the ethical principles is focused on
interventions at the early input stages but very few tools or methods during the middle building
and testing phases (Morley et al., 2019). Also, these conversations are frequently bound to the
academic community and related discourses, making practitioner access to these conversations
difficult. Comprehensive ethics education is critical to ensure that future generations of
practitioners, designers and engineers take their role as creators of futures seriously. We still face
a significant gap in background and understanding between, on the one hand, people from the
humanities and social sciences, and, on the other hand, people from the natural and engineering
sciences, both within and outside academia.

The Algorithmic Public Opinion: a Policy Overview 72

 Box 9 — The Future of Personalization Algorithms

Since algorithms are continuously being developed, it is possible that problems caused by
(news) personalization could become even more threatening in the future than is currently the
case. A focus on the potential development of personalization algorithms, therefore, becomes
essential. If we imagine the future of personalization, we can quite confidently argue that
algorithms – especially smart assistants such as Alexa or Siri – will guide us and make ever more
sophisticated and reliable decisions for us and about us. It is often imagined that algorithms will
wake up us softly with our most preferred music, make the perfect breakfast, as our physician
recommended, choose the street we take, the people we meet, the books and news we read,
similarly to nowadays, but even suggesting important choices of life – what to study, whether to
accept a job offer or whether to marry.

These kind of cognitive outsourcing are provided by AI-driven algorithmic decision-making and
decision-guidance. They promise us to enhance our lives preserving our time and energy and
nudging us towards healthier behaviors and better decisions. A problematic issue with the future
of personalization algorithms is that they can increasingly gratify us to the extent that we may
come to accept its pervasive and at times deceptive role as benevolent. This argument has been
discussed in terms of “psychological hedonism” (Gal, 2017): if personalized systems will become
‘pleasure machines’ able to predict our choices and simply grant them to us, will we be willing
to give up our autonomy? And, if so, under which conditions? Emerging markets and research
fields like the Internet of Things (IoT), cognitive and affective computing (cognitive science and
psychology) will play a significant role in the future of these algorithms. Clearly, there is a fair
amount of unpredictability in communication technology development, preventing precise
predictions regarding what future implementations of personalization algorithms will look like.
Yet, it is expected the rising of Ambient Intelligence in conjunction with the Internet of Things.
This construct offers a vision in which automatic smart online and offline environments and
devices interact with each other, taking an unprecedented number of decisions for us and about
us to cater to our inferred preferences, representing a new paradigm in the construction of
knowledge (Hildebrandt and Koops, 2010). Similarly, a related promise is the predictive ability of
algorithms, especially of emotion. Several emotion-decoding technologies are being developed
nowadays throughout all the potential inputs: from the face through the analysis of facial
expression, micro-gestures, micro-movements of muscles and eye tracking to other inputs such
as voice tones, body language, keyboard typing recognition etc. These, however, may still suffer
from the significant theoretical limitations highlighted (see Box 1 — The Reductionism of Profiling
Technologies).

The Algorithmic Public Opinion: a Policy Overview 73

 Box 10 — The Role of Messagging App

A significant challenge is represented by encrypted messaging apps, particularly for content

moderation and disinformation. For example, Whatsapp, Telegram and Signal can be understood
as social media insofar as content sharing among small and large groups, public communication,
interpersonal connection, and commercial transactions converge in key features of the app.
These platforms gained immense popularity in key markets like India, Brazil, and Indonesia.
Such social media’s shift towards more private, integrated, and encrypted services opens up new
challenges not only for content moderation on the app, but also for the regulation of platforms by
governments.

Despite encryption, WhatsApp moderates content both at the account and content level (Gillespie
et al., 2020). At the account level, the company uses machine learning to detect abusive behaviour,
disables over two million accounts per month, and scans unencrypted information such as
profile pictures, which has been instrumental in detecting child pornography activity within the
app. At the content level, accusations of disinformation and mob violence pushed WhatsApp
to implement measures to curb the virality of problematic content. Unregulated virality on the
app depends on a combination of affordances: encryption, groups of up to 256 people, and the
forward function. Most of the groups on the platform are family and friends, local community
and neighbourhood groups of less than ten people. In these intimate spaces, one might think that
content moderation would not be needed at all. However, in countries like Brazil, Malaysia and
India, some WhatsApp groups can be much larger and act as semi-public forums. The sharing of
news and the discussion of politics are popular; this communication takes place among strangers,
and context collapse may occur. The combination of large groups and users’ unlimited ability
to forward messages helped information on WhatsApp be easily shared at scale, potentially
encouraging the spread of misinformation.

The Algorithmic Public Opinion: a Policy Overview 74