Recommendation on

the Ethics
of Artificial
Intelligence
Adopted on 23 November 2021
Published in 2022 by the United Nations Educational, Scientific and Cultural Organization,
7, place de Fontenoy, 75352 Paris 07 SP, France

© UNESCO 2022

SHS/BIO/PI/2021/1

All rights reserved.

This publication is available in Open Access under the Attribution-NonCommercial-

ShareAlike 3.0 IGO (CC-BY-NC-SA 3.0 IGO) license (http://creativecommons.org/licenses/
by-nc-sa/3.0/igo/). By using the content of this publication, the users accept to be bound
by the terms of use of the UNESCO Open Access Repository (www.unesco.org/open-
access/terms-use-ccbyncsa-en).

Cover photo:
Irina Bg/Shutterstock.com; Artistdesign29/Shutterstock.com; kynny/gettyimages;
gorodenkoffy/gettyimages; Gorodenkoff/Shutterstock.com; dieddin/Shutterstock.com;
Gorodenkoff/Shutterstock.com; PopTika/Shutterstock.com; Horth Rasur/Shutterstock.com.

Inside photo:
p.8: ESB Professional/Shutterstock.com
p.11: Ruslana Iurchenko/Shutterstock.com
p.12: metamorworks/Shutterstock.com
p.14/15: Alexander Supertramp/Shutterstock.com
p.16: Wazzkii/Shutterstock.com
p.19: Mukesh Kumar Jwala/Shutterstock.com
p.21: supparsorn/Shutterstock.com
p.24: everything possible/Shutterstock.com
p.29: Gorodenkoff/Shutterstock.com
p.31: only_kim/Shutterstock.com
p.33: SeventyFour/Shutterstock.com
p.34: ESB Professional/Shutterstock.com
p.36: KlingSup/Shutterstock.com
p.38: Miriam Doerr Martin Frommherz/Shutterstock.com

Graphic design: Aurélia Mazoyer

Printed by UNESCO

Printed in France
Recommendation on
the Ethics
of Artificial
Intelligence
Adopted on 23 November 2021
Table of contents

Preamble 5

I. Scope of application 9

II. Aims and objectives 13

III. Values and principles 17

III.1 VALUES 18
III.2 PRINCIPLES 20

IV. Areas of policy action 25

POLICY AREA 1: ETHICAL IMPACT ASSESSMENT 26
POLICY AREA 2: ETHICAL GOVERNANCE AND STEWARDSHIP 27
POLICY AREA 3: DATA POLICY 29
POLICY AREA 4: DEVELOPMENT AND INTERNATIONAL COOPERATION 30
POLICY AREA 5: ENVIRONMENT AND ECOSYSTEMS 30
POLICY AREA 6: GENDER 32
POLICY AREA 7: CULTURE 32
POLICY AREA 8: EDUCATION AND RESEARCH 33
POLICY AREA 9: COMMUNICATION AND INFORMATION 35
POLICY AREA 10: ECONOMY AND LABOUR 36
POLICY AREA 11: HEALTH AND SOCIAL WELL-BEING 37

V. Monitoring and evaluation 39

VI. Utilization and exploitation of the present Recommendation 41

VII. Promotion of the present Recommendation 42

VIII. Final provisions 43

3
Preamble

The General Conference of the United Nations Educational, Also recognizing that AI technologies can deepen
Scientific and Cultural Organization (UNESCO), meeting in existing divides and inequalities in the world, within and
Paris from 9 to 24 November 2021, at its 41st session, between countries, and that justice, trust and fairness must be
upheld so that no country and no one should be left behind,
Recognizing the profound and dynamic positive and either by having fair access to AI technologies and enjoying
negative impacts of artificial intelligence (AI) on societies, their benefits or in the protection against their negative
environment, ecosystems and human lives, including the implications, while recognizing the different circumstances of
human mind, in part because of the new ways in which its use different countries and respecting the desire of some people
influences human thinking, interaction and decision-making not to take part in all technological developments,
and affects education, human, social and natural sciences,
culture, and communication and information, Conscious of the fact that all countries are facing an
acceleration in the use of information and communication
Recalling that, by the terms of its Constitution, UNESCO technologies and AI technologies, as well as an increasing
seeks to contribute to peace and security by promoting need for media and information literacy, and that the
collaboration among nations through education, the sciences, digital economy presents important societal, economic and
culture, and communication and information, in order to environmental challenges and opportunities of benefit-
further universal respect for justice, for the rule of law and sharing, especially for low- and middle-income countries
for the human rights and fundamental freedoms which are (LMICs), including but not limited to least developed countries
affirmed for the peoples of the world, (LDCs), landlocked developing countries (LLDCs) and small
Convinced that the Recommendation presented here, as island developing States (SIDS), requiring the recognition,
a standard-setting instrument developed through a global protection and promotion of endogenous cultures, values and
approach, based on international law, focusing on human knowledge in order to develop sustainable digital economies,
dignity and human rights, as well as gender equality, social Further recognizing that AI technologies have the
and economic justice and development, physical and mental potential to be beneficial to the environment and ecosystems,
well-being, diversity, interconnectedness, inclusiveness, and in order for those benefits to be realized, potential harms
and environmental and ecosystem protection can guide AI to and negative impacts on the environment and ecosystems
technologies in a responsible direction, should not be ignored but instead addressed,
Guided by the purposes and principles of the Charter of the Noting that addressing risks and ethical concerns should not
United Nations, hamper innovation and development but rather provide new
Considering that AI technologies can be of great service opportunities and stimulate ethically-conducted research and
to humanity and all countries can benefit from them, but also innovation that anchor AI technologies in human rights and
raise fundamental ethical concerns, for instance regarding fundamental freedoms, values and principles, and moral and
the biases they can embed and exacerbate, potentially ethical reflection,
resulting in discrimination, inequality, digital divides, exclusion Also recalling that in November 2019, the General
and a threat to cultural, social and biological diversity and Conference of UNESCO, at its 40th session, adopted 40 C/
social or economic divides; the need for transparency and Resolution 37, by which it mandated the Director-General
understandability of the workings of algorithms and the data “to prepare an international standard-setting instrument
with which they have been trained; and their potential impact on the ethics of artificial intelligence (AI) in the form of a
on, including but not limited to, human dignity, human rights recommendation”, which is to be submitted to the General
and fundamental freedoms, gender equality, democracy, Conference at its 41st session in 2021,
social, economic, political and cultural processes, scientific and
engineering practices, animal welfare, and the environment Recognizing that the development of AI technologies
and ecosystems, necessitates a commensurate increase in data, media and
information literacy as well as access to independent,

5
 pluralistic, trusted sources of information, including as part of of Indigenous Peoples (2007); the United Nations General
efforts to mitigate risks of misinformation, disinformation and Assembly resolution on the review of the World Summit on the
hate speech, and harm caused through the misuse of personal Information Society (A/RES/70/125) (2015); the United Nations
data, General Assembly Resolution on Transforming our world:
the 2030 Agenda for Sustainable Development (A/RES/70/1)
Observing that a normative framework for AI technologies (2015); the Recommendation Concerning the Preservation
and its social implications finds its basis in international and of, and Access to, Documentary Heritage Including in Digital
national legal frameworks, human rights and fundamental Form (2015); the Declaration of Ethical Principles in relation to
freedoms, ethics, need for access to data, information Climate Change (2017); the Recommendation on Science and
and knowledge, the freedom of research and innovation, Scientific Researchers (2017); the Internet Universality Indicators
human and environmental and ecosystem well-being, and (endorsed by UNESCO’s International Programme for the
connects ethical values and principles to the challenges and Development of Communication in 2018), including the ROAM
opportunities linked to AI technologies, based on common principles (endorsed by UNESCO’s General Conference in 2015);
understanding and shared aims, the Human Rights Council’s resolution on “The right to privacy
Also recognizing that ethical values and principles can in the digital age” (A/HRC/RES/42/15) (2019); and the Human
help develop and implement rights-based policy measures Rights Council’s resolution on “New and emerging digital
and legal norms, by providing guidance with a view to the fast technologies and human rights” (A/HRC/RES/41/11) (2019),
pace of technological development, Emphasizing that specific attention must be paid to
Also convinced that globally accepted ethical standards LMICs, including but not limited to LDCs, LLDCs and SIDS, as
for AI technologies, in full respect of international law, in they have their own capacity but have been underrepresented
particular human rights law, can play a key role in developing in the AI ethics debate, which raises concerns about neglecting
AI-related norms across the globe, local knowledge, cultural pluralism, value systems and the
demands of global fairness to deal with the positive and
Bearing in mind the Universal Declaration of Human negative impacts of AI technologies,
Rights (1948), the instruments of the international human
rights framework, including the Convention Relating to the Also conscious of the many existing national policies,
Status of Refugees (1951), the Discrimination (Employment and other frameworks and initiatives elaborated by relevant United
Occupation) Convention (1958), the International Convention Nations entities, intergovernmental organizations, including
on the Elimination of All Forms of Racial Discrimination (1965), regional organizations, as well as those by the private sector,
the International Covenant on Civil and Political Rights (1966), professional organizations, non-governmental organizations,
the International Covenant on Economic, Social and Cultural and the scientific community, related to the ethics and
Rights (1966), the Convention on the Elimination of All Forms regulation of AI technologies,
of Discrimination against Women (1979), the Convention on Further convinced that AI technologies can bring
the Rights of the Child (1989), and the Convention on the important benefits, but that achieving them can also amplify
Rights of Persons with Disabilities (2006), the Convention tension around innovation, asymmetric access to knowledge
against Discrimination in Education (1960), the Convention and technologies, including the digital and civic literacy deficit
on the Protection and Promotion of the Diversity of Cultural that limits the public’s ability to engage in topics related to
Expressions (2005), as well as any other relevant international AI, as well as barriers to access to information and gaps in
instruments, recommendations and declarations, capacity, human and institutional capacities, barriers to access
Also noting the United Nations Declaration on the Right to to technological innovation, and a lack of adequate physical
Development (1986); the Declaration on the Responsibilities and digital infrastructure and regulatory frameworks, including
of the Present Generations Towards Future Generations those related to data, all of which need to be addressed,
(1997); the Universal Declaration on Bioethics and Human
Rights (2005); the United Nations Declaration on the Rights

6
Underlining that the strengthening of global cooperation
and solidarity, including through multilateralism, is needed
to facilitate fair access to AI technologies and address the
challenges that they bring to diversity and interconnectivity of
cultures and ethical systems, to mitigate potential misuse, to
realize the full potential that AI can bring, especially in the area
of development, and to ensure that national AI strategies are
guided by ethical principles,

Taking fully into account that the rapid development

of AI technologies challenges their ethical implementation
and governance, as well as the respect for and protection of
cultural diversity, and has the potential to disrupt local and
regional ethical standards and values,

1. Adopts the present Recommendation on the Ethics

of Artificial Intelligence on this twenty-third day of
November 2021;

2. Recommends that Member States apply on a

voluntary basis the provisions of this Recommendation by
taking appropriate steps, including whatever legislative
or other measures may be required, in conformity with
the constitutional practice and governing structures
of each State, to give effect within their jurisdictions to
the principles and norms of the Recommendation in
conformity with international law, including international
human rights law;

3. Also recommends that Member States engage

all stakeholders, including business enterprises, to
ensure that they play their respective roles in the
implementation of this Recommendation; and
bring the Recommendation to the attention of the
authorities, bodies, research and academic organizations,
institutions and organizations in public, private and civil
society sectors involved in AI technologies, so that the
development and use of AI technologies are guided by
both sound scientific research as well as ethical analysis
and evaluation.

7
8
 I.
Scope of
application

9
 1. This Recommendation addresses ethical issues related to to range from research, design and development
the domain of Artificial Intelligence to the extent that they to deployment and use, including maintenance,
are within UNESCO’s mandate. It approaches AI ethics as operation, trade, financing, monitoring and
a systematic normative reflection, based on a holistic, evaluation, validation, end-of-use, disassembly and
comprehensive, multicultural and evolving framework termination. In addition, AI actors can be defined
of interdependent values, principles and actions that can as any actor involved in at least one stage of the AI
guide societies in dealing responsibly with the known system life cycle, and can refer both to natural and
and unknown impacts of AI technologies on human legal persons, such as researchers, programmers,
beings, societies and the environment and ecosystems, engineers, data scientists, end-users, business
and offers them a basis to accept or reject AI technologies. enterprises, universities and public and private
It considers ethics as a dynamic basis for the normative entities, among others.
evaluation and guidance of AI technologies, referring to
human dignity, well-being and the prevention of harm (c) AI systems raise new types of ethical issues that
as a compass and as rooted in the ethics of science and include, but are not limited to, their impact on
technology. decision-making, employment and labour, social
interaction, health care, education, media, access
2. This Recommendation does not have the ambition to to information, digital divide, personal data and
provide one single definition of AI, since such a definition consumer protection, environment, democracy,
would need to change over time, in accordance with rule of law, security and policing, dual use,
technological developments. Rather, its ambition is to and human rights and fundamental freedoms,
address those features of AI systems that are of central including freedom of expression, privacy and non-
ethical relevance. Therefore, this Recommendation discrimination. Furthermore, new ethical challenges
approaches AI systems as systems which have the are created by the potential of AI algorithms to
capacity to process data and information in a way that reproduce and reinforce existing biases, and thus to
resembles intelligent behaviour, and typically includes exacerbate already existing forms of discrimination,
aspects of reasoning, learning, perception, prediction, prejudice and stereotyping. Some of these issues
planning or control. Three elements have a central place are related to the capacity of AI systems to perform
in this approach: tasks which previously only living beings could
do, and which were in some cases even limited to
(a) AI systems are information-processing technologies human beings only. These characteristics give AI
that integrate models and algorithms that produce systems a profound, new role in human practices
a capacity to learn and to perform cognitive and society, as well as in their relationship with
tasks leading to outcomes such as prediction the environment and ecosystems, creating a new
and decision-making in material and virtual context for children and young people to grow
environments. AI systems are designed to operate up in, develop an understanding of the world
with varying degrees of autonomy by means of and themselves, critically understand media and
knowledge modelling and representation and by information, and learn to make decisions. In the
exploiting data and calculating correlations. AI long term, AI systems could challenge humans’
systems may include several methods, such as but special sense of experience and agency, raising
not limited to: additional concerns about, inter alia, human self-
(i) machine learning, including deep learning understanding, social, cultural and environmental
and reinforcement learning; interaction, autonomy, agency, worth and dignity.

(ii) machine reasoning, including planning, 3. This Recommendation pays specific attention to the
scheduling, knowledge representation and broader ethical implications of AI systems in relation
reasoning, search, and optimization. to the central domains of UNESCO: education, science,
culture, and communication and information, as explored
AI systems can be used in cyber-physical systems, in the 2019 Preliminary Study on the Ethics of Artificial
including the Internet of things, robotic systems, Intelligence by the UNESCO World Commission on Ethics
social robotics, and human-computer interfaces, of Scientific Knowledge and Technology (COMEST):
which involve control, perception, the processing
of data collected by sensors, and the operation of (a) Education, because living in digitalizing societies
actuators in the environment in which AI systems requires new educational practices, ethical
work. reflection, critical thinking, responsible design
practices and new skills, given the implications
(b) Ethical questions regarding AI systems pertain to all for the labour market, employability and civic
stages of the AI system life cycle, understood here participation.

10
(b) Science, in the broadest sense and including structuring and provision of information; the issues
all academic fields from the natural sciences of automated journalism and the algorithmic
and medical sciences to the social sciences and provision of news and moderation and curation of
humanities, as AI technologies bring new research content on social media and search engines are just
capacities and approaches, have implications a few examples raising issues related to access to
for our concepts of scientific understanding and information, disinformation, misinformation, hate
explanation, and create a new basis for decision- speech, the emergence of new forms of societal
making. narratives, discrimination, freedom of expression,
privacy and media and information literacy, among
(c) Cultural identity and diversity, as AI technologies others.
can enrich cultural and creative industries, but can
also lead to an increased concentration of supply of 4. This Recommendation is addressed to Member States,
cultural content, data, markets and income in the both as AI actors and as authorities responsible for
hands of only a few actors, with potential negative developing legal and regulatory frameworks throughout
implications for the diversity and pluralism of the entire AI system life cycle, and for promoting business
languages, media, cultural expressions, participation responsibility. It also provides ethical guidance to all AI
and equality. actors, including the public and private sectors, by
providing a basis for an ethical impact assessment of AI
(d) Communication and information, as AI technologies systems throughout their life cycle.
play an increasingly important role in the processing,

11
12
 II.
Aims and
objectives

13
 5. This Recommendation aims to provide a basis to make 7. Because the complexity of the ethical issues
AI systems work for the good of humanity, individuals, surrounding AI necessitates the cooperation of multiple
societies and the environment and ecosystems, and to stakeholders across the various levels and sectors of
prevent harm. It also aims at stimulating the peaceful use international, regional and national communities, this
of AI systems. Recommendation aims to enable stakeholders to take
shared responsibility based on a global and intercultural
6. In addition to the existing ethical frameworks regarding dialogue.
AI around the world, this Recommendation aims to bring
a globally accepted normative instrument that focuses 8. The objectives of this Recommendation are:
not only on the articulation of values and principles,
but also on their practical realization, via concrete (a) to provide a universal framework of values, principles
policy recommendations, with a strong emphasis on and actions to guide States in the formulation
inclusion issues of gender equality and protection of the of their legislation, policies or other instruments
environment and ecosystems. regarding AI, consistent with international law;

14
(b) to guide the actions of individuals, groups, (d) to foster multi-stakeholder, multidisciplinary and
communities, institutions and private sector pluralistic dialogue and consensus building about
companies to ensure the embedding of ethics in all ethical issues relating to AI systems;
stages of the AI system life cycle;
(e) to promote equitable access to developments and
(c) to protect, promote and respect human rights knowledge in the field of AI and the sharing of
and fundamental freedoms, human dignity and benefits, with particular attention to the needs and
equality, including gender equality; to safeguard contributions of LMICs, including LDCs, LLDCs and
the interests of present and future generations; SIDS.
to preserve the environment, biodiversity and
ecosystems; and to respect cultural diversity in all
stages of the AI system life cycle;

15
16
 III.
Values and
principles

17
 9. The values and principles included below should be human rights and fundamental freedoms. In all cases, any
respected by all actors in the AI system life cycle, in the first possible limitations on human rights and fundamental
place and, where needed and appropriate, be promoted freedoms must have a lawful basis, and be reasonable,
through amendments to the existing and elaboration necessary and proportionate, and consistent with States’
of new legislation, regulations and business guidelines. obligations under international law. To navigate such
This must comply with international law, including the scenarios judiciously will typically require engagement
United Nations Charter and Member States’ human rights with a broad range of appropriate stakeholders, making
obligations, and should be in line with internationally use of social dialogue, as well as ethical deliberation, due
agreed social, political, environmental, educational, diligence and impact assessment.
scientific and economic sustainability objectives, such
as the United Nations Sustainable Development Goals 12. The trustworthiness and integrity of the life cycle of AI
(SDGs). systems is essential to ensure that AI technologies will
work for the good of humanity, individuals, societies
10. Values play a powerful role as motivating ideals in shaping and the environment and ecosystems, and embody the
policy measures and legal norms. While the set of values values and principles set out in this Recommendation.
outlined below thus inspires desirable behaviour and People should have good reason to trust that AI systems
represents the foundations of principles, the principles can bring individual and shared benefits, while adequate
unpack the values underlying them more concretely so measures are taken to mitigate risks. An essential
that the values can be more easily operationalized in requirement for trustworthiness is that, throughout their
policy statements and actions. life cycle, AI systems are subject to thorough monitoring
by the relevant stakeholders as appropriate. As
11. While all the values and principles outlined below are trustworthiness is an outcome of the operationalization
desirable per se, in any practical contexts, there may be of the principles in this document, the policy actions
tensions between these values and principles. In any proposed in this Recommendation are all directed at
given situation, a contextual assessment will be necessary promoting trustworthiness in all stages of the AI system
to manage potential tensions, taking into account the life cycle.
principle of proportionality and in compliance with

III.1 VALUES

Respect, protection and promotion of human life” should be left open to individuals or groups, as long
rights and fundamental freedoms and human as there is no violation or abuse of human rights and
dignity fundamental freedoms, or the dignity of humans in terms
of this definition.
13. The inviolable and inherent dignity of every human
constitutes the foundation for the universal, indivisible, 15. Persons may interact with AI systems throughout their
inalienable, interdependent and interrelated system of life cycle and receive assistance from them, such as
human rights and fundamental freedoms. Therefore, care for vulnerable people or people in vulnerable
respect, protection and promotion of human dignity situations, including but not limited to children, older
and rights as established by international law, including persons, persons with disabilities or the ill. Within such
international human rights law, is essential throughout interactions, persons should never be objectified, nor
the life cycle of AI systems. Human dignity relates to should their dignity be otherwise undermined, or human
the recognition of the intrinsic and equal worth of rights and fundamental freedoms violated or abused.
each individual human being, regardless of race, colour,
descent, gender, age, language, religion, political opinion, 16. Human rights and fundamental freedoms must be
national origin, ethnic origin, social origin, economic respected, protected and promoted throughout the
or social condition of birth, or disability and any other life cycle of AI systems. Governments, private sector,
grounds. civil society, international organizations, technical
communities and academia must respect human rights
14. No human being or human community should be harmed instruments and frameworks in their interventions in the
or subordinated, whether physically, economically, processes surrounding the life cycle of AI systems. New
socially, politically, culturally or mentally during any technologies need to provide new means to advocate,
phase of the life cycle of AI systems. Throughout the life defend and exercise human rights and not to infringe
cycle of AI systems, the quality of life of human beings them.
should be enhanced, while the definition of “quality of

18
Environment and ecosystem flourishing race, colour, descent, gender, age, language, religion,
political opinion, national origin, ethnic origin, social
17. Environmental and ecosystem flourishing should be origin, economic or social condition of birth, or disability
recognized, protected and promoted through the life and any other grounds.
cycle of AI systems. Furthermore, environment and
ecosystems are the existential necessity for humanity 20. The scope of lifestyle choices, beliefs, opinions,
and other living beings to be able to enjoy the benefits expressions or personal experiences, including the
of advances in AI. optional use of AI systems and the co-design of these
architectures should not be restricted during any phase
18. All actors involved in the life cycle of AI systems must of the life cycle of AI systems.
comply with applicable international law and domestic
legislation, standards and practices, such as precaution, 21. Furthermore, efforts, including international cooperation,
designed for environmental and ecosystem protection should be made to overcome, and never take advantage
and restoration, and sustainable development. They of, the lack of necessary technological infrastructure,
should reduce the environmental impact of AI systems, education and skills, as well as legal frameworks,
including but not limited to its carbon footprint, to ensure particularly in LMICs, LDCs, LLDCs and SIDS, affecting
the minimization of climate change and environmental communities.
risk factors, and prevent the unsustainable exploitation,
use and transformation of natural resources contributing
to the deterioration of the environment and the Living in peaceful, just and interconnected
societies
degradation of ecosystems.
22. AI actors should play a participative and enabling role
to ensure peaceful and just societies, which is based on
Ensuring diversity and inclusiveness
an interconnected future for the benefit of all, consistent
19. Respect, protection and promotion of diversity and with human rights and fundamental freedoms. The
inclusiveness should be ensured throughout the life cycle value of living in peaceful and just societies points to the
of AI systems, consistent with international law, including potential of AI systems to contribute throughout their
human rights law. This may be done by promoting active life cycle to the interconnectedness of all living creatures
participation of all individuals or groups regardless of with each other and with the natural environment.

19
 23. The notion of humans being interconnected is based 24. This value demands that peace, inclusiveness and
on the knowledge that every human belongs to a justice, equity and interconnectedness should be
greater whole, which thrives when all its constituent promoted throughout the life cycle of AI systems, in
parts are enabled to thrive. Living in peaceful, just and so far as the processes of the life cycle of AI systems
interconnected societies requires an organic, immediate, should not segregate, objectify or undermine freedom
uncalculated bond of solidarity, characterized by a and autonomous decision-making as well as the safety
permanent search for peaceful relations, tending of human beings and communities, divide and turn
towards care for others and the natural environment in individuals and groups against each other, or threaten
the broadest sense of the term. the coexistence between humans, other living beings
and the natural environment.

III.2 PRINCIPLES

Proportionality and Do No Harm Fairness and non-discrimination

25. It should be recognized that AI technologies do not 28. AI actors should promote social justice and safeguard
necessarily, per se, ensure human and environmental fairness and non-discrimination of any kind in compliance
and ecosystem flourishing. Furthermore, none of the with international law. This implies an inclusive approach
processes related to the AI system life cycle shall exceed to ensuring that the benefits of AI technologies are
what is necessary to achieve legitimate aims or objectives available and accessible to all, taking into consideration
and should be appropriate to the context. In the event the specific needs of different age groups, cultural systems,
of possible occurrence of any harm to human beings, different language groups, persons with disabilities,
human rights and fundamental freedoms, communities girls and women, and disadvantaged, marginalized and
and society at large or the environment and ecosystems, vulnerable people or people in vulnerable situations.
the implementation of procedures for risk assessment Member States should work to promote inclusive access
and the adoption of measures in order to preclude the for all, including local communities, to AI systems with
occurrence of such harm should be ensured. locally relevant content and services, and with respect
for multilingualism and cultural diversity. Member States
26. The choice to use AI systems and which AI method to should work to tackle digital divides and ensure inclusive
use should be justified in the following ways: (a) the AI access to and participation in the development of AI. At
method chosen should be appropriate and proportional the national level, Member States should promote equity
to achieve a given legitimate aim; (b) the AI method between rural and urban areas, and among all persons
chosen should not infringe upon the foundational values regardless of race, colour, descent, gender, age, language,
captured in this document, in particular, its use must not religion, political opinion, national origin, ethnic origin,
violate or abuse human rights; and (c) the AI method social origin, economic or social condition of birth, or
should be appropriate to the context and should be disability and any other grounds, in terms of access
based on rigorous scientific foundations. In scenarios to and participation in the AI system life cycle. At the
where decisions are understood to have an impact that international level, the most technologically advanced
is irreversible or difficult to reverse or may involve life countries have a responsibility of solidarity with the least
and death decisions, final human determination should advanced to ensure that the benefits of AI technologies
apply. In particular, AI systems should not be used for are shared such that access to and participation in the
social scoring or mass surveillance purposes. AI system life cycle for the latter contributes to a fairer
world order with regard to information, communication,
Safety and security culture, education, research and socio-economic and
political stability.
27. Unwanted harms (safety risks), as well as vulnerabilities
to attack (security risks) should be avoided and should 29. AI actors should make all reasonable efforts to minimize
be addressed, prevented and eliminated throughout the and avoid reinforcing or perpetuating discriminatory
life cycle of AI systems to ensure human, environmental or biased applications and outcomes throughout the
and ecosystem safety and security. Safe and secure AI will life cycle of the AI system to ensure fairness of such
be enabled by the development of sustainable, privacy- systems. Effective remedy should be available against
protective data access frameworks that foster better discrimination and biased algorithmic determination.
training and validation of AI models utilizing quality data.

20
30. Furthermore, digital and knowledge divides within and Right to Privacy, and Data Protection
between countries need to be addressed throughout
an AI system life cycle, including in terms of access and 32. Privacy, a right essential to the protection of human
quality of access to technology and data, in accordance dignity, human autonomy and human agency, must
with relevant national, regional and international be respected, protected and promoted throughout the
legal frameworks, as well as in terms of connectivity, life cycle of AI systems. It is important that data for AI
knowledge and skills and meaningful participation of the systems be collected, used, shared, archived and deleted
affected communities, such that every person is treated in ways that are consistent with international law and
equitably. in line with the values and principles set forth in this
Recommendation, while respecting relevant national,
regional and international legal frameworks.
Sustainability
33. Adequate data protection frameworks and governance
31. The development of sustainable societies relies on mechanisms should be established in a multi-stakeholder
the achievement of a complex set of objectives on a approach at the national or international level, protected
continuum of human, social, cultural, economic and by judicial systems, and ensured throughout the life
environmental dimensions. The advent of AI technologies cycle of AI systems. Data protection frameworks and
can either benefit sustainability objectives or hinder any related mechanisms should take reference from
their realization, depending on how they are applied international data protection principles and standards
across countries with varying levels of development. The concerning the collection, use and disclosure of personal
continuous assessment of the human, social, cultural, data and exercise of their rights by data subjects while
economic and environmental impact of AI technologies ensuring a legitimate aim and a valid legal basis for the
should therefore be carried out with full cognizance of processing of personal data, including informed consent.
the implications of AI technologies for sustainability
as a set of constantly evolving goals across a range of 34. Algorithmic systems require adequate privacy impact
dimensions, such as currently identified in the Sustainable assessments, which also include societal and ethical
Development Goals (SDGs) of the United Nations. considerations of their use and an innovative use of the

21
 privacy by design approach. AI actors need to ensure that sector company or public sector institution able to
they are accountable for the design and implementation review and correct the decision. AI actors should inform
of AI systems in such a way as to ensure that personal users when a product or service is provided directly or
information is protected throughout the life cycle of the with the assistance of AI systems in a proper and timely
AI system. manner.

39. From a socio-technical lens, greater transparency

Human oversight and determination contributes to more peaceful, just, democratic and
inclusive societies. It allows for public scrutiny that can
35. Member States should ensure that it is always possible to decrease corruption and discrimination, and can also
attribute ethical and legal responsibility for any stage of help detect and prevent negative impacts on human
the life cycle of AI systems, as well as in cases of remedy rights. Transparency aims at providing appropriate
related to AI systems, to physical persons or to existing information to the respective addressees to enable their
legal entities. Human oversight refers thus not only understanding and foster trust. Specific to the AI system,
to individual human oversight, but to inclusive public transparency can enable people to understand how
oversight, as appropriate. each stage of an AI system is put in place, appropriate
36. It may be the case that sometimes humans would to the context and sensitivity of the AI system. It may
choose to rely on AI systems for reasons of efficacy, but also include insight into factors that affect a specific
the decision to cede control in limited contexts remains prediction or decision, and whether or not appropriate
that of humans, as humans can resort to AI systems in assurances (such as safety or fairness measures) are in
decision-making and acting, but an AI system can never place. In cases of serious threats of adverse human rights
replace ultimate human responsibility and accountability. impacts, transparency may also require the sharing of
As a rule, life and death decisions should not be ceded to code or datasets.
AI systems. 40. Explainability refers to making intelligible and providing
insight into the outcome of AI systems. The explainability
Transparency and explainability of AI systems also refers to the understandability of the
input, output and the functioning of each algorithmic
37. The transparency and explainability of AI systems are building block and how it contributes to the outcome
often essential preconditions to ensure the respect, of the systems. Thus, explainability is closely related to
protection and promotion of human rights, fundamental transparency, as outcomes and sub-processes leading
freedoms and ethical principles. Transparency is to outcomes should aim to be understandable and
necessary for relevant national and international liability traceable, appropriate to the context. AI actors should
regimes to work effectively. A lack of transparency could commit to ensuring that the algorithms developed are
also undermine the possibility of effectively challenging explainable. In the case of AI applications that impact the
decisions based on outcomes produced by AI systems end user in a way that is not temporary, easily reversible
and may thereby infringe the right to a fair trial and or otherwise low risk, it should be ensured that the
effective remedy, and limits the areas in which these meaningful explanation is provided with any decision
systems can be legally used. that resulted in the action taken in order for the outcome
to be considered transparent.
38. While efforts need to be made to increase transparency
and explainability of AI systems, including those with 41. Transparency and explainability relate closely to adequate
extra-territorial impact, throughout their life cycle to responsibility and accountability measures, as well as to
support democratic governance, the level of transparency the trustworthiness of AI systems.
and explainability should always be appropriate to the
context and impact, as there may be a need to balance
between transparency and explainability and other Responsibility and accountability
principles such as privacy, safety and security. People 42. AI actors and Member States should respect, protect and
should be fully informed when a decision is informed promote human rights and fundamental freedoms, and
by or is made on the basis of AI algorithms, including should also promote the protection of the environment
when it affects their safety or human rights, and in those and ecosystems, assuming their respective ethical and
circumstances should have the opportunity to request legal responsibility, in accordance with national and
explanatory information from the relevant AI actor or international law, in particular Member States’ human
public sector institutions. In addition, individuals should rights obligations, and ethical guidance throughout
be able to access the reasons for a decision affecting the life cycle of AI systems, including with respect to AI
their rights and freedoms, and have the option of making actors within their effective territory and control. The
submissions to a designated staff member of the private ethical responsibility and liability for the decisions and

22
 actions based in any way on an AI system should always their impact on human rights and access to rights, as well
ultimately be attributable to AI actors corresponding to as on the environment and ecosystems.
their role in the life cycle of the AI system.

43. Appropriate oversight, impact assessment, audit and Multi-stakeholder and adaptive governance and
due diligence mechanisms, including whistle-blowers’ collaboration
protection, should be developed to ensure accountability
46. International law and national sovereignty must be
for AI systems and their impact throughout their life
respected in the use of data. That means that States,
cycle. Both technical and institutional designs should
complying with international law, can regulate the data
ensure auditability and traceability of (the working of ) AI
generated within or passing through their territories,
systems in particular to address any conflicts with human
and take measures towards effective regulation of data,
rights norms and standards and threats to environmental
including data protection, based on respect for the right
and ecosystem well-being.
to privacy in accordance with international law and other
human rights norms and standards.
Awareness and literacy
47. Participation of different stakeholders throughout
44. Public awareness and understanding of AI technologies the AI system life cycle is necessary for inclusive
and the value of data should be promoted through open approaches to AI governance, enabling the benefits
and accessible education, civic engagement, digital to be shared by all, and to contribute to sustainable
skills and AI ethics training, media and information development. Stakeholders include but are not limited
literacy and training led jointly by governments, to governments, intergovernmental organizations,
intergovernmental organizations, civil society, academia, the technical community, civil society, researchers and
the media, community leaders and the private sector, academia, media, education, policy-makers, private
and considering the existing linguistic, social and cultural sector companies, human rights institutions and equality
diversity, to ensure effective public participation so that bodies, anti-discrimination monitoring bodies, and
all members of society can take informed decisions groups for youth and children. The adoption of open
about their use of AI systems and be protected from standards and interoperability to facilitate collaboration
undue influence. should be in place. Measures should be adopted to take
into account shifts in technologies, the emergence of
45. Learning about the impact of AI systems should include new groups of stakeholders, and to allow for meaningful
learning about, through and for human rights and participation by marginalized groups, communities and
fundamental freedoms, meaning that the approach and individuals and, where relevant, in the case of Indigenous
understanding of AI systems should be grounded by Peoples, respect for the self-governance of their data.

23
24
 IV.
Areas of
policy action

25
 48. The policy actions described in the following policy 49. UNESCO recognizes that Member States will be
areas operationalize the values and principles set out in at different stages of readiness to implement this
this Recommendation. The main action is for Member Recommendation, in terms of scientific, technological,
States to put in place effective measures, including, for economic, educational, legal, regulatory, infrastructural,
example, policy frameworks or mechanisms, and to societal, cultural and other dimensions. It is noted that
ensure that other stakeholders, such as private sector “readiness” here is a dynamic status. In order to enable
companies, academic and research institutions, and the effective implementation of this Recommendation,
civil society adhere to them by, among other actions, UNESCO will therefore: (1) develop a readiness assessment
encouraging all stakeholders to develop human rights, methodology to assist interested Member States in
rule of law, democracy, and ethical impact assessment identifying their status at specific moments of their
and due diligence tools in line with guidance including readiness trajectory along a continuum of dimensions;
the United Nations Guiding Principles on Business and and (2) ensure support for interested Member States in
Human Rights. The process for developing such policies terms of developing a UNESCO methodology for Ethical
or mechanisms should be inclusive of all stakeholders Impact Assessment (EIA) of AI technologies, sharing
and should take into account the circumstances and of best practices, assessment guidelines and other
priorities of each Member State. UNESCO can be a partner mechanisms and analytical work.
and support Member States in the development as well
as monitoring and evaluation of policy mechanisms.

POLICY AREA 1: ETHICAL IMPACT ASSESSMENT

50. Member States should introduce frameworks for impact including in real-world conditions if needed, as part of
assessments, such as ethical impact assessment, to the Ethical Impact Assessment, before releasing them in
identify and assess benefits, concerns and risks of AI the market.
systems, as well as appropriate risk prevention, mitigation
and monitoring measures, among other assurance 52. Member States and business enterprises should
mechanisms. Such impact assessments should identify implement appropriate measures to monitor all phases
impacts on human rights and fundamental freedoms, in of an AI system life cycle, including the functioning of
particular but not limited to the rights of marginalized algorithms used for decision-making, the data, as well
and vulnerable people or people in vulnerable situations, as AI actors involved in the process, especially in public
labour rights, the environment and ecosystems and services and where direct end-user interaction is needed,
ethical and social implications, and facilitate citizen as part of ethical impact assessment. Member States’
participation in line with the values and principles set human rights law obligations should form part of the
forth in this Recommendation. ethical aspects of AI system assessments.

51. Member States and private sector companies should 53. Governments should adopt a regulatory framework that
develop due diligence and oversight mechanisms sets out a procedure, particularly for public authorities,
to identify, prevent, mitigate and account for how to carry out ethical impact assessments on AI systems
they address the impact of AI systems on the respect to predict consequences, mitigate risks, avoid harmful
for human rights, rule of law and inclusive societies. consequences, facilitate citizen participation and
Member States should also be able to assess the socio- address societal challenges. The assessment should also
economic impact of AI systems on poverty and ensure establish appropriate oversight mechanisms, including
that the gap between people living in wealth and auditability, traceability and explainability, which enable
poverty, as well as the digital divide among and within the assessment of algorithms, data and design processes,
countries, are not increased with the massive adoption as well as include external review of AI systems. Ethical
of AI technologies at present and in the future. In impact assessments should be transparent and open to
order to do this, in particular, enforceable transparency the public, where appropriate. Such assessments should
protocols should be implemented, corresponding to the also be multidisciplinary, multi-stakeholder, multicultural,
access to information, including information of public pluralistic and inclusive. The public authorities should be
interest held by private entities. Member States, private required to monitor the AI systems implemented and/or
sector companies and civil society should investigate deployed by those authorities by introducing appropriate
the sociological and psychological effects of AI-based mechanisms and tools.
recommendations on humans in their decision-making
autonomy. AI systems identified as potential risks to
human rights should be broadly tested by AI actors,

26
POLICY AREA 2: ETHICAL GOVERNANCE AND STEWARDSHIP
54. Member States should ensure that AI governance 58. Member States should encourage public entities, private
mechanisms are inclusive, transparent, multidisciplinary, sector companies and civil society organizations to
multilateral (this includes the possibility of mitigation and involve different stakeholders in their AI governance
redress of harm across borders) and multi-stakeholder. and to consider adding the role of an independent AI
In particular, governance should include aspects of Ethics Officer or some other mechanism to oversee
anticipation, and effective protection, monitoring of ethical impact assessment, auditing and continuous
impact, enforcement and redress. monitoring efforts and ensure ethical guidance of AI
systems. Member States, private sector companies and
55. Member States should ensure that harms caused through civil society organizations, with the support of UNESCO,
AI systems are investigated and redressed, by enacting are encouraged to create a network of independent AI
strong enforcement mechanisms and remedial actions, Ethics Officers to give support to this process at national,
to make certain that human rights and fundamental regional and international levels.
freedoms and the rule of law are respected in the digital
world and in the physical world. Such mechanisms 59. Member States should foster the development of, and
and actions should include remediation mechanisms access to, a digital ecosystem for ethical and inclusive
provided by private and public sector companies. The development of AI systems at the national level, including
auditability and traceability of AI systems should be to address gaps in access to the AI system life cycle, while
promoted to this end. In addition, Member States should contributing to international collaboration. Such an
strengthen their institutional capacities to deliver on this ecosystem includes, in particular, digital technologies
commitment and should collaborate with researchers and infrastructure, and mechanisms for sharing AI
and other stakeholders to investigate, prevent and knowledge, as appropriate.
mitigate any potentially malicious uses of AI systems.
60. Member States should establish mechanisms, in
56. Member States are encouraged to develop national collaboration with international organizations,
and regional AI strategies and to consider forms of soft transnational corporations, academic institutions and
governance such as a certification mechanism for AI civil society, to ensure the active participation of all
systems and the mutual recognition of their certification, Member States, especially LMICs, in particular LDCs,
according to the sensitivity of the application domain LLDCs and SIDS, in international discussions concerning
and expected impact on human rights, the environment AI governance. This can be through the provision of
and ecosystems, and other ethical considerations set funds, ensuring equal regional participation, or any
forth in this Recommendation. Such a mechanism other mechanisms. Furthermore, in order to ensure the
might include different levels of audit of systems, data, inclusiveness of AI fora, Member States should facilitate
and adherence to ethical guidelines and to procedural the travel of AI actors in and out of their territory,
requirements in view of ethical aspects. At the same especially from LMICs, in particular LDCs, LLDCs and SIDS,
time, such a mechanism should not hinder innovation or for the purpose of participating in these fora.
disadvantage small and medium enterprises or start-ups,
civil society as well as research and science organizations, 61. Amendments to the existing or elaboration of new
as a result of an excessive administrative burden. These national legislation addressing AI systems must comply
mechanisms should also include a regular monitoring with Member States’ human rights law obligations and
component to ensure system robustness and continued promote human rights and fundamental freedoms
integrity and adherence to ethical guidelines over the throughout the AI system life cycle. Promotion thereof
entire life cycle of the AI system, requiring re-certification should also take the form of governance initiatives,
if necessary. good exemplars of collaborative practices regarding AI
systems, and national and international technical and
57. Member States and public authorities should carry out methodological guidelines as AI technologies advance.
transparent self-assessment of existing and proposed Diverse sectors, including the private sector, in their
AI systems, which, in particular, should include the practices regarding AI systems must respect, protect and
assessment of whether the adoption of AI is appropriate promote human rights and fundamental freedoms using
and, if so, should include further assessment to determine existing and new instruments in combination with this
what the appropriate method is, as well as assessment as Recommendation.
to whether such adoption would result in violations or
abuses of Member States’ human rights law obligations, 62. Member States that acquire Al systems for human rights-
and if that is the case, prohibit its use. sensitive use cases, such as law enforcement, welfare,
employment, media and information providers, health
care and the independent judiciary system should provide

27
 mechanisms to monitor the social and economic impact 67. Member States should implement policies to promote
of such systems by appropriate oversight authorities, and increase diversity and inclusiveness that reflect
including independent data protection authorities, their populations in AI development teams and training
sectoral oversight and public bodies responsible for datasets, and to ensure equal access to AI technologies
oversight. and their benefits, particularly for marginalized groups,
both from rural and urban zones.
63. Member States should enhance the capacity of the
judiciary to make decisions related to AI systems as per 68. Member States should develop, review and adapt,
the rule of law and in line with international law and as appropriate, regulatory frameworks to achieve
standards, including in the use of AI systems in their accountability and responsibility for the content and
deliberations, while ensuring that the principle of human outcomes of AI systems at the different phases of their life
oversight is upheld. In case AI systems are used by the cycle. Member States should, where necessary, introduce
judiciary, sufficient safeguards are needed to guarantee liability frameworks or clarify the interpretation of existing
inter alia the protection of fundamental human rights, the frameworks to ensure the attribution of accountability
rule of law, judicial independence as well as the principle for the outcomes and the functioning of AI systems.
of human oversight, and to ensure a trustworthy, public Furthermore, when developing regulatory frameworks,
interest-oriented and human-centric development and Member States should, in particular, take into account that
use of AI systems in the judiciary. ultimate responsibility and accountability must always lie
with natural or legal persons and that AI systems should
64. Member States should ensure that governments and not be given legal personality themselves. To ensure
multilateral organizations play a leading role in ensuring this, such regulatory frameworks should be consistent
the safety and security of AI systems, with multi- with the principle of human oversight and establish a
stakeholder participation. Specifically, Member States, comprehensive approach focused on AI actors and the
international organizations and other relevant bodies technological processes involved across the different
should develop international standards that describe stages of the AI system life cycle.
measurable, testable levels of safety and transparency,
so that systems can be objectively assessed and levels 69. In order to establish norms where these do not exist,
of compliance determined. Furthermore, Member States or to adapt the existing legal frameworks, Member
and business enterprises should continuously support States should involve all AI actors (including, but not
strategic research on potential safety and security risks limited to, researchers, representatives of civil society
of AI technologies and should encourage research into and law enforcement, insurers, investors, manufacturers,
transparency and explainability, inclusion and literacy by engineers, lawyers and users). The norms can mature into
putting additional funding into those areas for different best practices, laws and regulations. Member States are
domains and at different levels, such as technical and further encouraged to use mechanisms such as policy
natural language. prototypes and regulatory sandboxes to accelerate
the development of laws, regulations and policies,
65. Member States should implement policies to ensure that including regular reviews thereof, in line with the rapid
the actions of AI actors are consistent with international development of new technologies and ensure that laws
human rights law, standards and principles throughout and regulations can be tested in a safe environment
the life cycle of AI systems, while taking into full before being officially adopted. Member States should
consideration the current cultural and social diversities, support local governments in the development of local
including local customs and religious traditions, with policies, regulations and laws in line with national and
due regard to the precedence and universality of human international legal frameworks.
rights.
70. Member States should set clear requirements for AI
66. Member States should put in place mechanisms to system transparency and explainability so as to help
require AI actors to disclose and combat any kind of ensure the trustworthiness of the full AI system life
stereotyping in the outcomes of AI systems and data, cycle. Such requirements should involve the design and
whether by design or by negligence, and to ensure that implementation of impact mechanisms that take into
training data sets for AI systems do not foster cultural, consideration the nature of application domain, intended
economic or social inequalities, prejudice, the spreading use, target audience and feasibility of each particular AI
of disinformation and misinformation, and disruption system.
of freedom of expression and access to information.
Particular attention should be given to regions where the
data are scarce.

28
POLICY AREA 3: DATA POLICY
71. Member States should work to develop data governance appropriate safeguards for the processing of sensitive
strategies that ensure the continual evaluation of the data; an appropriate level of data protection; effective and
quality of training data for AI systems including the meaningful accountability schemes and mechanisms;
adequacy of the data collection and selection processes, the full enjoyment of the data subjects’ rights and the
proper data security and protection measures, as well as ability to access and erase their personal data in AI
feedback mechanisms to learn from mistakes and share systems, except for certain circumstances in compliance
best practices among all AI actors. with international law; an appropriate level of protection
in full compliance with data protection legislation where
72. Member States should put in place appropriate safeguards data are being used for commercial purposes such as
to protect the right to privacy in accordance with enabling micro-targeted advertising, transferred cross-
international law, including addressing concerns such border; and an effective independent oversight as part of
as surveillance. Member States should, among others, a data governance mechanism which keeps individuals
adopt or enforce legislative frameworks that provide in control of their personal data and fosters the benefits
appropriate protection, compliant with international of a free flow of information internationally, including
law. Member States should strongly encourage all AI access to data.
actors, including business enterprises, to follow existing
international standards and, in particular, to carry out 74. Member States should establish their data policies or
adequate privacy impact assessments, as part of ethical equivalent frameworks, or reinforce existing ones, to
impact assessments, which take into account the wider ensure full security for personal data and sensitive data,
socio-economic impact of the intended data processing, which, if disclosed, may cause exceptional damage, injury
and to apply privacy by design in their systems. or hardship to individuals. Examples include data relating
Privacy should be respected, protected and promoted to offences, criminal proceedings and convictions, and
throughout the life cycle of AI systems. related security measures; biometric, genetic and health
data; and -personal data such as that relating to race,
73. Member States should ensure that individuals retain colour, descent, gender, age, language, religion, political
rights over their personal data and are protected by opinion, national origin, ethnic origin, social origin,
a framework, which notably foresees: transparency;

29
 economic or social condition of birth, or disability and basis, including consent of data subjects, when required
any other characteristics. by law. Standards for annotating datasets should be
encouraged, including disaggregating data on gender
75. Member States should promote open data. In this regard, and other bases, so it can easily be determined how a
Member States should consider reviewing their policies dataset is gathered and what properties it has.
and regulatory frameworks, including on access to
information and open government to reflect AI-specific 77. Member States, as also suggested in the report of
requirements and promoting mechanisms, such as open the United Nations Secretary-General’s High-level
repositories for publicly funded or publicly held data and Panel on Digital Cooperation, with the support of the
source code and data trusts, to support the safe, fair, legal United Nations and UNESCO, should adopt a digital
and ethical sharing of data, among others. commons approach to data where appropriate, increase
interoperability of tools and datasets and interfaces of
76. Member States should promote and facilitate the use systems hosting data, and encourage private sector
of quality and robust datasets for training, development companies to share the data they collect with all
and use of AI systems, and exercise vigilance in stakeholders, as appropriate, for research, innovation or
overseeing their collection and use. This could, if possible public benefits. They should also promote public and
and feasible, include investing in the creation of gold private efforts to create collaborative platforms to share
standard datasets, including open and trustworthy quality data in trusted and secured data spaces.
datasets, which are diverse, constructed on a valid legal

POLICY AREA 4: DEVELOPMENT AND INTERNATIONAL COOPERATION

78. Member States and transnational corporations research and innovation centres and networks that
should prioritize AI ethics by including discussions of promote greater participation and leadership of
AI-related ethical issues into relevant international, researchers from LMICs and other countries, including
intergovernmental and multi-stakeholder fora. LDCs, LLDCs and SIDS.

79. Member States should ensure that the use of AI in areas 82. Member States should promote AI ethics research by
of development such as education, science, culture, engaging international organizations and research
communication and information, health care, agriculture institutions, as well as transnational corporations, that can
and food supply, environment, natural resource and be a basis for the ethical use of AI systems by public and
infrastructure management, economic planning and private entities, including research into the applicability
growth, among others, adheres to the values and of specific ethical frameworks in specific cultures and
principles set forth in this Recommendation. contexts, and the possibilities to develop technologically
feasible solutions in line with these frameworks.
80. Member States should work through international
organizations to provide platforms for international 83. Member States should encourage international
cooperation on AI for development, including by cooperation and collaboration in the field of AI to bridge
contributing expertise, funding, data, domain knowledge, geo-technological lines. Technological exchanges and
infrastructure, and facilitating multi-stakeholder consultations should take place between Member States
collaboration to tackle challenging development and their populations, between the public and private
problems, especially for LMICs, in particular LDCs, LLDCs sectors, and between and among the most and least
and SIDS. technologically advanced countries in full respect of
international law.
81. Member States should work to promote international
collaboration on AI research and innovation, including

POLICY AREA 5: ENVIRONMENT AND ECOSYSTEMS

84. Member States and business enterprises should for supporting the manufacturing of AI technologies,
assess the direct and indirect environmental impact and reduce the environmental impact of AI systems
throughout the AI system life cycle, including, but not and data infrastructures. Member States should ensure
limited to, its carbon footprint, energy consumption and compliance of all AI actors with environmental law,
the environmental impact of raw material extraction policies and practices.

30
85. Member States should introduce incentives, when (e) Enable and promote the mainstreaming of
needed and appropriate, to ensure the development sustainable infrastructure, sustainable business
and adoption of rights-based and ethical AI-powered models and sustainable finance for sustainable
solutions for disaster risk resilience; the monitoring, development.
protection and regeneration of the environment and
ecosystems; and the preservation of the planet. These (f ) Detect pollutants or predict levels of pollution and
AI systems should involve the participation of local thus help relevant stakeholders identify, plan and
and indigenous communities throughout the life cycle put in place targeted interventions to prevent and
of AI systems and should support circular economy reduce pollution and exposure.
type approaches and sustainable consumption and 86. When choosing AI methods, given the potential data-
production patterns. Some examples include using AI intensive or resource-intensive character of some of
systems, when needed and appropriate, to: them and the respective impact on the environment,
(a) Support the protection, monitoring and Member States should ensure that AI actors, in line with
management of natural resources. the principle of proportionality, favour data, energy and
resource-efficient AI methods. Requirements should
(b) Support the prediction, prevention, control and be developed to ensure that appropriate evidence is
mitigation of climate-related problems. available to show that an AI application will have the
intended effect, or that safeguards accompanying an
(c) Support a more efficient and sustainable food AI application can support the justification for its use.
ecosystem. If this cannot be done, the precautionary principle
(d) Support the acceleration of access to and mass must be favoured, and in instances where there are
adoption of sustainable energy. disproportionate negative impacts on the environment,
AI should not be used.

31
 POLICY AREA 6: GENDER
87. Member States should ensure that the potential for digital 90. Member States should ensure that gender stereotyping
technologies and artificial intelligence to contribute to and discriminatory biases are not translated into AI
achieving gender equality is fully maximized, and must systems, and instead identify and proactively redress
ensure that the human rights and fundamental freedoms these. Efforts are necessary to avoid the compounding
of girls and women, and their safety and integrity are not negative effect of technological divides in achieving
violated at any stage of the AI system life cycle. Moreover, gender equality and avoiding violence such as
Ethical Impact Assessment should include a transversal harassment, bullying or trafficking of girls and women
gender perspective. and under-represented groups, including in the online
domain.
88. Member States should have dedicated funds from their
public budgets linked to financing gender-responsive 91. Member States should encourage female entrepreneurship,
schemes, ensure that national digital policies include a participation and engagement in all stages of an AI system
gender action plan, and develop relevant policies, for life cycle by offering and promoting economic, regulatory
example, on labour education, targeted at supporting incentives, among other incentives and support schemes, as
girls and women to make sure they are not left out of well as policies that aim at a balanced gender participation
the digital economy powered by AI. Special investment in AI research in academia, gender representation on digital
in providing targeted programmes and gender- and AI companies’ top management positions, boards of
specific language, to increase the opportunities of directors and research teams. Member States should ensure
girls’ and women’s participation in science, technology, that public funds (for innovation, research and technologies)
engineering, and mathematics (STEM), including are channelled to inclusive programmes and companies,
information and communication technologies (ICT) with clear gender representation, and that private funds are
disciplines, preparedness, employability, equal career similarly encouraged through affirmative action principles.
development and professional growth of girls and Policies on harassment-free environments should be
women, should be considered and implemented. developed and enforced, together with the encouragement
of the transfer of best practices on how to promote diversity
89. Member States should ensure that the potential of AI throughout the AI system life cycle.
systems to advance the achievement of gender equality
is realized. They should ensure that these technologies 92. Member States should promote gender diversity
do not exacerbate the already wide gender gaps in AI research in academia and industry by offering
existing in several fields in the analogue world, and incentives to girls and women to enter the field, putting
instead eliminate those gaps. These gaps include: the in place mechanisms to fight gender stereotyping and
gender wage gap; the unequal representation in certain harassment within the AI research community, and
professions and activities; the lack of representation encouraging academic and private entities to share best
at top management positions, boards of directors, or practices on how to enhance gender diversity.
research teams in the AI field; the education gap; the
digital and AI access, adoption, usage and affordability 93. UNESCO can help form a repository of best practices
gap; and the unequal distribution of unpaid work and of for incentivizing the participation of girls, women and
the caring responsibilities in our societies. under-represented groups in all stages of the AI system
life cycle.

POLICY AREA 7: CULTURE

94. Member States are encouraged to incorporate AI 95. Member States are encouraged to examine and
systems, where appropriate, in the preservation, address the cultural impact of AI systems, especially
enrichment, understanding, promotion, management natural language processing (NLP) applications such
and accessibility of tangible, documentary and intangible as automated translation and voice assistants, on the
cultural heritage, including endangered languages as well nuances of human language and expression. Such
as indigenous languages and knowledges, for example assessments should provide input for the design and
by introducing or updating educational programmes implementation of strategies that maximize the benefits
related to the application of AI systems in these areas, from these systems by bridging cultural gaps and
where appropriate, and by ensuring a participatory increasing human understanding, as well as addressing
approach, targeted at institutions and the public. the negative implications such as the reduction of use,
which could lead to the disappearance of endangered

32
 languages, local dialects, and tonal and cultural variations 98. Member States should engage technology companies
associated with human language and expression. and other stakeholders to promote a diverse supply of
and plural access to cultural expressions, and in particular
96. Member States should promote AI education and digital to ensure that algorithmic recommendation enhances
training for artists and creative professionals to assess the the visibility and discoverability of local content.
suitability of AI technologies for use in their profession,
and contribute to the design and implementation of 99. Member States should foster new research at the
suitable AI technologies, as AI technologies are being intersection between AI and intellectual property (IP), for
used to create, produce, distribute, broadcast and example to determine whether or how to protect with
consume a variety of cultural goods and services, bearing IP rights the works created by means of Al technologies.
in mind the importance of preserving cultural heritage, Member States should also assess how AI technologies
diversity and artistic freedom. are affecting the rights or interests of IP owners, whose
works are used to research, develop, train or implement
97. Member States should promote awareness and AI applications.
evaluation of AI tools among local cultural industries and
small and medium enterprises working in the field of 100. Member States should encourage museums, galleries,
culture, to avoid the risk of concentration in the cultural libraries and archives at the national level to use AI
market. systems to highlight their collections and enhance their
libraries, databases and knowledge base, while also
providing access to their users.

POLICY AREA 8: EDUCATION AND RESEARCH

101. Member States should work with international 102. Member States should promote the acquisition of
organizations, educational institutions and private “prerequisite skills” for AI education, such as basic
and non-governmental entities to provide adequate literacy, numeracy, coding and digital skills, and media
AI literacy education to the public on all levels in all and information literacy, as well as critical and creative
countries in order to empower people and reduce the thinking, teamwork, communication, socio-emotional
digital divides and digital access inequalities resulting and AI ethics skills, especially in countries and in regions
from the wide adoption of AI systems. or areas within countries where there are notable gaps in
the education of these skills.

33
 103. Member States should promote general awareness should be considered when discussing the adoption of
programmes about AI developments, including on AI technologies in education. AI systems used in learning
data and the opportunities and challenges brought should be subject to strict requirements when it comes
about by AI technologies, the impact of AI systems on to the monitoring, assessment of abilities, or prediction
human rights and their implications, including children’s of the learners’ behaviours. AI should support the
rights. These programmes should be accessible to non- learning process without reducing cognitive abilities and
technical as well as technical groups. without extracting sensitive information, in compliance
with relevant personal data protection standards. The
104. Member States should encourage research initiatives data handed over to acquire knowledge collected
on the responsible and ethical use of AI technologies during the learner’s interactions with the AI system must
in teaching, teacher training and e-learning, among not be subject to misuse, misappropriation or criminal
other issues, to enhance opportunities and mitigate the exploitation, including for commercial purposes.
challenges and risks involved in this area. The initiatives
should be accompanied by an adequate assessment of 105. Member States should promote the participation and
the quality of education and impact on students and leadership of girls and women, diverse ethnicities and
teachers of the use of AI technologies. Member States cultures, persons with disabilities, marginalized and
should also ensure that AI technologies empower vulnerable people or people in vulnerable situations,
students and teachers and enhance their experience, minorities and all persons not enjoying the full benefits of
bearing in mind that relational and social aspects and digital inclusion, in AI education programmes at all levels,
the value of traditional forms of education are vital in as well as the monitoring and sharing of best practices in
teacher-student and student-student relationships and this regard with other Member States.

34
106. Member States should develop, in accordance with their 109. Member States should encourage private sector
national education programmes and traditions, AI ethics companies to facilitate the access of the scientific
curricula for all levels, and promote cross-collaboration community to their data for research, especially in
between AI technical skills education and humanistic, LMICs, in particular LDCs, LLDCs and SIDS. This access
ethical and social aspects of AI education. Online courses should conform to relevant privacy and data protection
and digital resources of AI ethics education should be standards.
developed in local languages, including indigenous
languages, and take into account the diversity of 110. To ensure a critical evaluation of AI research and
environments, especially ensuring accessibility of formats proper monitoring of potential misuses or adverse
for persons with disabilities. effects, Member States should ensure that any future
developments with regards to AI technologies should be
107. Member States should promote and support AI research, based on rigorous and independent scientific research,
notably AI ethics research, including for example through and promote interdisciplinary AI research by including
investing in such research or by creating incentives for disciplines other than science, technology, engineering
the public and private sectors to invest in this area, and mathematics (STEM), such as cultural studies,
recognizing that research contributes significantly education, ethics, international relations, law, linguistics,
to the further development and improvement of AI philosophy, political science, sociology and psychology.
technologies with a view to promoting international
law and the values and principles set forth in this 111. Recognizing that AI technologies present great
Recommendation. Member States should also publicly opportunities to help advance scientific knowledge
promote the best practices of, and cooperation with, and practice, especially in traditionally model-driven
researchers and companies who develop AI in an ethical disciplines, Member States should encourage scientific
manner. communities to be aware of the benefits, limits and
risks of their use; this includes attempting to ensure that
108. Member States should ensure that AI researchers are conclusions drawn from data-driven approaches, models
trained in research ethics and require them to include and treatments are robust and sound. Furthermore,
ethical considerations in their designs, products and Member States should welcome and support the role of
publications, especially in the analyses of the datasets the scientific community in contributing to policy and in
they use, how they are annotated, and the quality and cultivating awareness of the strengths and weaknesses of
scope of the results with possible applications. AI technologies.

POLICY AREA 9: COMMUNICATION AND INFORMATION

112. Member States should use AI systems to improve access of content, and appeal mechanisms that allow users to
to information and knowledge. This can include support seek redress.
to researchers, academia, journalists, the general public
and developers, to enhance freedom of expression, 114. Member States should invest in and promote digital and
academic and scientific freedoms, access to information, media and information literacy skills to strengthen critical
and increased proactive disclosure of official data and thinking and competencies needed to understand the
information. use and implication of AI systems, in order to mitigate and
counter disinformation, misinformation and hate speech.
113. Member States should ensure that AI actors respect A better understanding and evaluation of both the
and promote freedom of expression as well as access positive and potentially harmful effects of recommender
to information with regard to automated content systems should be part of those efforts.
generation, moderation and curation. Appropriate
frameworks, including regulation, should enable 115. Member States should create enabling environments
transparency of online communication and information for media to have the rights and resources to effectively
operators and ensure users have access to a diversity of report on the benefits and harms of AI systems, and also
viewpoints, as well as processes for prompt notification to encourage media to make ethical use of AI systems in
the users on the reasons for removal or other treatment their operations.

35
 POLICY AREA 10: ECONOMY AND LABOUR
116. Member States should assess and address the impact stakeholders, including workers and unions to ensure a
of AI systems on labour markets and its implications for fair transition for at-risk employees. This includes putting
education requirements, in all countries and with special in place upskilling and reskilling programmes, finding
emphasis on countries where the economy is labour- effective mechanisms of retaining employees during
intensive. This can include the introduction of a wider those transition periods, and exploring “safety net”
range of “core” and interdisciplinary skills at all education programmes for those who cannot be retrained. Member
levels to provide current workers and new generations a States should develop and implement programmes
fair chance of finding jobs in a rapidly changing market, to research and address the challenges identified that
and to ensure their awareness of the ethical aspects could include upskilling and reskilling, enhanced social
of AI systems. Skills such as “learning how to learn”, protection, proactive industry policies and interventions,
communication, critical thinking, teamwork, empathy, tax benefits, new taxation forms, among others. Member
and the ability to transfer one’s knowledge across States should ensure that there is sufficient public funding
domains, should be taught alongside specialist, technical to support these programmes. Relevant regulations,
skills, as well as low-skilled tasks. Being transparent about such as tax regimes, should be carefully examined and
what skills are in demand and updating curricula around changed if needed to counteract the consequences of
these are key. unemployment caused by AI-based automation.

117. Member States should support collaboration agreements 119. Member States should encourage and support
among governments, academic institutions, vocational researchers to analyse the impact of AI systems on the
education and training institutions, industry, workers’ local labour environment in order to anticipate future
organizations and civil society to bridge the gap of skillset trends and challenges. These studies should have an
requirements to align training programmes and strategies interdisciplinary approach and investigate the impact of
with the implications of the future of work and the needs AI systems on economic, social and geographic sectors,
of industry, including small and medium enterprises. as well as on human-robot interactions and human-
Project-based teaching and learning approaches for AI human relationships, in order to advise on reskilling and
should be promoted, allowing for partnerships between redeployment best practices.
public institutions, private sector companies, universities
and research centres. 120. Member States should take appropriate steps to
ensure competitive markets and consumer protection,
118. Member States should work with private sector considering possible measures and mechanisms at
companies, civil society organizations and other national, regional and international levels, to prevent

36
 abuse of dominant market positions, including by human capacity and regulations, among other factors. AI
monopolies, in relation to AI systems throughout their actors developing AI systems in countries which have
life cycle, whether these are data, research, technology, established or adopted ethical standards on AI should
or market. Member States should prevent the resulting respect these standards when exporting these products,
inequalities, assess relevant markets and promote developing or applying their AI systems in countries
competitive markets. Due consideration should be given where such standards may not exist, while respecting
to LMICs, in particular LDCs, LLDCs and SIDS, which are applicable international law and domestic legislation,
more exposed and vulnerable to the possibility of abuses standards and practices of these countries.
of market dominance as a result of a lack of infrastructure,

POLICY AREA 11: HEALTH AND SOCIAL WELL-BEING

121. Member States should endeavour to employ effective (e) ensuring the human care and final decision of
AI systems for improving human health and protecting diagnosis and treatment are taken always by
the right to life, including mitigating disease outbreaks, humans while acknowledging that AI systems can
while building and maintaining international solidarity to also assist in their work;
tackle global health risks and uncertainties, and ensure
that their deployment of AI systems in health care be (f ) ensuring, where necessary, the review of AI systems
consistent with international law and their human rights by an ethical research committee prior to clinical
law obligations. Member States should ensure that actors use.
involved in health care AI systems take into consideration 124. Member States should establish research on the effects
the importance of a patient’s relationships with their and regulation of potential harms to mental health related
family and with health care staff. to AI systems, such as higher degrees of depression,
122. Member States should ensure that the development and anxiety, social isolation, developing addiction, trafficking,
deployment of AI systems related to health in general radicalization and misinformation, among others.
and mental health in particular, paying due attention to 125. Member States should develop guidelines for human-
children and youth, is regulated to the effect that they are robot interactions and their impact on human-human
safe, effective, efficient, scientifically and medically proven relationships, based on research and directed at the future
and enable evidence-based innovation and medical development of robots, and with special attention to the
progress. Moreover, in the related area of digital health mental and physical health of human beings. Particular
interventions, Member States are strongly encouraged to attention should be given to the use of robots in health
actively involve patients and their representatives in all care and the care for older persons and persons with
relevant steps of the development of the system. disabilities, in education, and robots for use by children,
123. Member States should pay particular attention in toy robots, chatbots and companion robots for children
regulating prediction, detection and treatment and adults. Furthermore, assistance of AI technologies
solutions for health care in AI applications by: should be applied to increase the safety and ergonomic
use of robots, including in a human-robot working
(a) ensuring oversight to minimize and mitigate bias; environment. Special attention should be paid to the
possibility of using AI to manipulate and abuse human
(b) ensuring that the professional, the patient, caregiver cognitive biases.
or service user is included as a “domain expert” in
the team in all relevant steps when developing the 126. Member States should ensure that human-robot
algorithms; interactions comply with the same values and principles
that apply to any other AI systems, including human
(c) paying due attention to privacy because of the rights and fundamental freedoms, the promotion of
potential need for being medically monitored and diversity, and the protection of vulnerable people or
ensuring that all relevant national and international people in vulnerable situations. Ethical questions related
data protection requirements are met; to AI-powered systems for neurotechnologies and brain-
(d) ensuring effective mechanisms so that those whose computer interfaces should be considered in order to
personal data is being analysed are aware of and preserve human dignity and autonomy.
provide informed consent for the use and analysis 127. Member States should ensure that users can easily
of their data, without preventing access to health identify whether they are interacting with a living
care; being, or with an AI system imitating human or animal

37
 characteristics, and can effectively refuse such interaction attention to the psychological and cognitive impact that
and request human intervention. these systems can have on children and young people.
This should be done using multiple norms, principles,
128. Member States should implement policies to raise protocols, disciplinary approaches, and assessment
awareness about the anthropomorphization of AI of the modification of behaviours and habits, as well
technologies and technologies that recognize and as careful evaluation of the downstream cultural and
mimic human emotions, including in the language societal impacts. Furthermore, Member States should
used to mention them, and assess the manifestations, encourage research on the effect of AI technologies on
ethical implications and possible limitations of such health system performance and health outcomes.
anthropomorphization, in particular in the context of
robot-human interaction and especially when children 130. Member States, as well as all stakeholders, should put in
are involved. place mechanisms to meaningfully engage children and
young people in conversations, debates and decision-
129. Member States should encourage and promote making with regard to the impact of AI systems on their
collaborative research into the effects of long-term lives and futures.
interaction of people with AI systems, paying particular

38
V.
Monitoring
and evaluation
131. Member States should, according to their specific (d) strengthening the research- and evidence-based
conditions, governing structures and constitutional analysis of and reporting on policies regarding AI
provisions, credibly and transparently monitor and ethics;
evaluate policies, programmes and mechanisms related
to ethics of AI, using a combination of quantitative and (e) collecting and disseminating progress, innovations,
qualitative approaches. To support Member States, research reports, scientific publications, data and
UNESCO can contribute by: statistics regarding policies for AI ethics, including
through existing initiatives, to support sharing best
(a) developing a UNESCO methodology for Ethical practices and mutual learning, and to advance the
Impact Assessment (EIA) of AI technologies based implementation of this Recommendation.
on rigorous scientific research and grounded in
international human rights law, guidance for its 132. Processes for monitoring and evaluation should ensure
implementation in all stages of the AI system life broad participation of all stakeholders, including, but
cycle, and capacity-building materials to support not limited to, vulnerable people or people in vulnerable
Member States’ efforts to train government officials, situations. Social, cultural and gender diversity should be
policy-makers and other relevant AI actors on EIA ensured, with a view to improving learning processes
methodology; and strengthening the connections between findings,
decision-making, transparency and accountability for
(b) developing a UNESCO readiness assessment results.
methodology to assist Member States in identifying
their status at specific moments of their readiness 133. In the interests of promoting best policies and practices
trajectory along a continuum of dimensions; related to ethics of AI, appropriate tools and indicators
should be developed for assessing the effectiveness and
(c) developing a UNESCO methodology to evaluate ex efficiency thereof against agreed standards, priorities and
ante and ex post the effectiveness and efficiency targets, including specific targets for persons belonging
of the policies for AI ethics and incentives against to disadvantaged, marginalized populations, and
defined objectives; vulnerable people or people in vulnerable situations, as
well as the impact of AI systems at individual and societal
levels. The monitoring and assessment of the impact of

39
 AI systems and related AI ethics policies and practices 134. In particular, Member States may wish to consider
should be carried out continuously in a systematic possible mechanisms for monitoring and evaluation,
way proportionate to the relevant risks. This should be such as an ethics commission, AI ethics observatory,
based on internationally agreed frameworks and involve repository covering human rights-compliant and ethical
evaluations of private and public institutions, providers development of AI systems, or contributions to existing
and programmes, including self-evaluations, as well as initiatives by addressing adherence to ethical principles
tracer studies and the development of sets of indicators. across UNESCO’s areas of competence, an experience-
Data collection and processing should be conducted in sharing mechanism, AI regulatory sandboxes, and an
accordance with international law, national legislation assessment guide for all AI actors to evaluate their
on data protection and data privacy, and the values and adherence to policy recommendations mentioned in
principles outlined in this Recommendation. this document.

40
VI.
Utilization and
exploitation
of the present
Recommendation

135. Member States and all other stakeholders as identified cooperating with all relevant national and international
in this Recommendation should respect, promote and governmental and non-governmental organizations,
protect the ethical values, principles and standards as well as transnational corporations and scientific
regarding AI that are identified in this Recommendation, organizations, whose activities fall within the scope and
and should take all feasible steps to give effect to its objectives of this Recommendation. The development of
policy recommendations. a UNESCO Ethical Impact Assessment methodology and
the establishment of national commissions for the ethics
136. Member States should strive to extend and complement of AI can be important instruments for this.
their own action in respect of this Recommendation, by

41
 VII.
Promotion of
the present
Recommendation
137. UNESCO has the vocation to be the principal United 139. Even though, within UNESCO, the mandate to promote
Nations agency to promote and disseminate this and protect falls within the authority of governments
Recommendation, and accordingly will work in and intergovernmental bodies, civil society will be an
collaboration with other relevant United Nations entities, important actor to advocate for the public sector’s
while respecting their mandate and avoiding duplication interests and therefore UNESCO needs to ensure and
of work. promote its legitimacy.

138. UNESCO, including its bodies, such as the World

Commission on the Ethics of Scientific Knowledge
and Technology (COMEST), the International Bioethics
Committee (IBC) and the Intergovernmental Bioethics
Committee (IGBC), will also work in collaboration
with other international, regional and sub-regional
governmental and non-governmental organizations.

42
VIII.
Final
provisions
140. This Recommendation needs to be understood as a
whole, and the foundational values and principles are to
be understood as complementary and interrelated.

141. Nothing in this Recommendation may be interpreted

as replacing, altering or otherwise prejudicing States’
obligations or rights under international law, or as
approval for any State, other political, economic or social
actor, group or person to engage in any activity or perform
any act contrary to human rights, fundamental freedoms,
human dignity and concern for the environment and
ecosystems, both living and non-living.

43
Social and Human Sciences Sector
7, place de Fontenoy
75352 Paris 07 SP France
[email protected]
on.unesco.org/EthicsAI

Follow us
@UNESCO #AI #HumanAI