Governance with teeth:

How human rights can strengthen FAT and

ethics initiatives on artificial intelligence

April 2019
First published by ARTICLE 19, 2019

ARTICLE 19
Free Word Centre
60 Farringdon Road
London EC1R 3GA
UK
www.article19.org
T: +44 20 7324 2500
E: info@article19.org
Tw: @article19org
Fb: facebook.com/article19org

ISBN: 978-1-910793-42-8

Text and analysis © ARTICLE 19, 2018 under Creative Commons Attribution-Non-
Commercial-ShareAlike 2.5 licence. To access the full legal text of this licence, please
visit: http://creativecommons.org/licenses/by-ncsa/2.5/legalcode.

ARTICLE 19 works for a world where all people everywhere can freely express
themselves and actively engage in public life without fear of discrimination. We do this
by working on two interlocking freedoms, which set the foundation for all our work.
The Freedom to Speak concerns everyone’s right to express and disseminate opinions,
ideas and information through any means, as well as to disagree from, and question
power-holders. The Freedom to Know concerns the right to demand and receive
information by power-holders for transparency good governance and sustainable
development. When either of these freedoms comes under threat, by the failure of
power-holders to adequately protect them, ARTICLE 19 speaks with one voice, through
courts of law, through global and regional organisations, and through civil society
wherever we are present.
Contents
About us 4

Executive Summary 6

I. Introduction 8

2. The societal impact of AI: two approaches 9

2.1 The normative approach: Ethics initiatives 9

2.2 The technical approach: Fairness, Accountability and Transparency (FAT) 13

3. Towards a human rights-based approach 17

3.1 A human rights-based approach to Ethics 19

3.2 A human rights-based approach to FAT 20

3.3 How might FAT and ethics initiatives help improve a

human rights-based approach? 21

4. Recommendations 22

Endnotes 23
About us
ARTICLE 19 is a global human rights organisation that protects and promotes the
right to freedom of expression and information around the world. Established in
1987 in London, ARTICLE 19 monitors threats to freedom of expression in different
regions of the world, and develops long-term strategies to address them.

ARTICLE 19 actively promotes human rights-respecting Artificial Intelligence (AI),

and investigates the impacts of algorithmic decision making on people’s lives.
Through policy engagement and research, our published work on AI and freedom
of expression thus far includes Privacy and Freedom of Expression in the Age
of Artificial Intelligence and a policy brief on algorithms and automated decision
making.

We have also provided expert input on related topics and served on multiple
committees dedicated to AI and human rights. We have served as invited experts
to the UNESCO World Commission on the Ethics of Scientific Knowledge and
Technology (COMEST) and several other UN processes, including consultations
with various Special Rapporteurs. We have made a submission to the UK House
of Lords Select Committee on AI; offered expert input to the Council of Europe
committee MSI-AUT; and made several submissions to the AI and ethics initiative of
the Institute of Electrical and Electronics Engineering (IEEE), where we also maintain
co-chairship of several working groups of the initiative. We hold membership in the
Partnership on AI and have given guidance on the development of AI for network
management at the Internet Engineering Task Force.

4
Executive Summary
As artificial intelligence (AI) is increasingly integrated into societies, its potential
impact on democracy and society has given rise to important debates about how AI
systems should be governed. Some stakeholders have put their focus on building
normative ethical principles, while others have gravitated towards a technical
discussion of how to build fair, accountable, and transparent AI systems. A third
approach has been to apply existing legal human rights frameworks to guide the
development of AI that is human rights-respecting through design, development and
deployment.

In this paper, ARTICLE 19 considers the ethical and technical approaches in the
field so far. We identify the contours and limitations of these parallel discussions,
and then propose a human rights-based approach to each of them. The intention
behind this paper is to explore how a human rights-based approach can
constructively inform and enhance these efforts and present key recommendations
to stakeholders:

We call on industry to:

1. Affirm their commitment to the UN Guiding Principles on business and

human rights.

2. Embed international human rights standards, particularly those of protecting

freedom of expression and privacy, in the development and deployment of AI
systems in a manner that lays out precise obligations and commitments for
companies building these technologies.

3. Explicitly invoke a human rights-based approach to ethical codes of conduct;

identify duty bearers and rights holders.

4. Publicly articulate what safeguards are in place to monitor and evaluate

ongoing ethical efforts, and institute accountability and redressal
mechanisms.

5. Enhance transparency around ethical AI initiatives, from the constitution

of ethical boards (including their powers and functions), to how ethical
principles are internally evaluated, to the impact of these initiatives on
businesses.

6. Ensure that AI systems undergo rigorous human rights impact assessments,

and create feedback and continuous auditing mechanisms for the same.

5
 7. As a continuation of the above, meaningfully engage with civil society and
academia at each stage of the process, to cultivate constructive criticism as
part of internal deliberation processes.

We call on states to:

1. Reaffirm their commitment to international human rights standards,

particularly those protecting freedom of expression and privacy, and let this
commitment guide their specific national AI strategies.

2. Ensure that AI systems in the public sector undergo adequate human rights
impact assessments, due diligence, and continuous auditing. These are not
systems that can simply be rolled out. They should instead be tailored to the
exact context and use for which they are intended.

3. Root the design, development, and deployment of AI systems in

constitutional guarantees and human rights standards.

4. Hold AI systems to accountability, responsibility, and constitutional

standards without dilution or exception.

5. Ensure that national and international efforts around AI are equally informed
by human rights concerns, constitutional standards, and the public interest
as they are by industry concerns.

6. Establish necessary infrastructure and resources to provide remedy for

human rights violations caused by AI systems.

We call on civil society to:

1. Work towards greater transparency and accountability of mechanisms

(such impact assessments, the Universal Periodic Review, and rights-based
mechanisms at national levels) and institutions (like the Human Rights
Council, General Assembly, Special Rapporteurs, national and regional
institutions that focus on compliance with and enforcement of human rights)
that make up a rights-based approach.

2. Advocate for the promotion and protection of human rights in the context of
AI systems in a manner that is also informed by technical considerations
and limitations.

3. Actively engage with ethical initiatives to integrate strong human rights

language and protections into their substantive and procedural functioning.

6
I. Introduction
As artificial intelligence (AI)1 has demonstrated its power to revolutionise
fundamental systems of communication, commerce, labor, and public services,
it has captured the attention of the technology industry, public officials, and civil
society. The potential of AI to perform tasks with speed and at scale beyond human
capability has fueled great excitement. AI systems are already deeply embedded
in our everyday lives - from helping us navigate through morning traffic to offering
up the day’s news, to more nefarious uses of systems for surveillance,2 warfare,3
and oppressing democratic dissent.4 Yet many of the most powerful stakeholders
in the field have only just begun to consider the impact of AI systems on society,
democracy, rights, and justice.

Technologists, researchers, companies and governments are grappling with

the challenges that AI poses for society, democracy, and governance, while
simultaneously trying to develop frameworks to guide this technological
development. Thus far, the solutions that different stakeholders have put forth to
tackle problematic uses of AI and reckon with their unintended consequences are
insufficient, as we will show in the pages that follow.

In these processes, at least three distinct approaches to AI standards and principles

have emerged. Some stakeholders have put their focus on building normative ethical
principles, while others have gravitated towards determining what fair, accountable,
transparent AI systems look like from a technical standpoint. A third approach is
working to apply existing legal human rights frameworks to guide the development
of AI. Our aim is to explore how these conversations can constructively inform one
another.

In this paper, we analyse two approaches to understanding and addressing the

societal impacts of AI: 1) a normative approach, with a focus on ethics; and 2) a
technical approach, with a focus on fairness, accountability, and transparency (FAT).
We identify the contours and limitations of these approaches, and then demonstrate
how a human rights-based approach can strengthen them. Finally our analysis
concludes with concrete recommendations to stakeholder groups.

7
2. The societal impact of AI: two
approaches
2.1 The normative approach: Ethics initiatives

AI systems raise myriad questions for society and democracy, only some of which
are covered5 or addressed by existing laws.6 In order to fill these perceived gaps, a
vocal group of governments, industry players, academics, and civil society actors
have been promoting principles or frameworks for ethical AI.7 While there is an
abstract awareness of what “ethics” generally means, there is no precise or shared
understanding of the term. It has been subject to multiple interpretations by various
stakeholders, and is often defined by industry actors who use it on a case-by-case
basis.

Proponents of this normative ethical approach believe that it enables stakeholders

to identify opportunities that are socially acceptable or preferable, while at the same
time potentially averting costly mistakes by elucidating what is socially unacceptable.
As Floridi et al write, “With an analogy, it is the difference between playing according
to the rules, and playing well, so that one may win the game.”8 Ethics initiatives allow
conversations to transcend a single jurisdiction, and encourage input from various
stakeholders to crystallise high level commitments to what AI should look like.

At the time of writing this paper, at least 25 countries have published national AI
strategies9 and ethical task forces have cropped up around the world. The European
Commission’s High Level Expert Group on AI has laid down “Ethics Guidelines for
Trustworthy AI” focusing on respect for human autonomy, prevention of harm,
fairness, and explicability.10 Companies are constituting ethical boards, publishing
ethical principles for AI, and taking part in multi-stakeholder initiatives in this space
as well. Technical organisations like the Association for Computing Machinery
(ACM) and the Institute for Electrical and Electronics Engineers (IEEE) have published
ethical principles for autonomous systems. And academic11 and civil society actors
have engaged via government consultations12 and multi-stakeholder forums like the
Partnership on AI (PAI).13

This plethora of initiatives underlines the need for a framework to discuss the
desired impact of new technologies on society - it pushes for an articulation of what
ought to be done. The current debate around AI and ethics is rich, multi-disciplinary,
and takes various forms, including ethics boards, principles, and public statements.
ARTICLE 19 recognises the importance of these efforts, but also believes that there
is more reason to be critical than accepting of ethics initiatives, as we will discuss in
the next section.

8
2.1.3 Critical gaps in ethics initiatives

A primary reason to be critical of ethics initiatives in isolation is that they, more often
than not, are not actionable. They do not afford mechanisms that lead to tangible
change. The various principles developed by industry and states have, as of yet,
failed to develop strong accompanying accountability mechanisms. They lack
concrete and narrowly defined language,14 independent oversight or enforcement
mechanisms, and clear transparency and reporting requirements. This means that
no matter how laudable the principles are, there is no way to hold governments
or companies to said principles. The general lack of transparency mechanisms
leaves no pathway for other stakeholders to know whether or not companies and
governments are complying with their own principles. And in cases where non-
compliance is revealed, there are inadequate mechanisms to hold companies and
governments accountable for their wrongdoing.

For instance, after Google received pushback from its own employees surrounding
Project Maven, a partnership with the US Department of Defence to improve drone
targeting using AI, the company published a set of AI principles that elucidated
its commitment to ethics, and made a public pledge to refrain from building
certain types of technology.15 But Google has not disclosed to what extent these
principles are embedded in concrete work in the company, and there has been no
demonstrable change in how the company has altered its internal decision making
processes. This is particularly worrying because in the case of public-private
partnerships such as Project Maven, the accountability that governments otherwise
owe the public is potentially diluted by the use of technology built behind closed
doors and vague, non-binding commitments.

This lack of accountability can even take a toll when AI systems are still being
developed. Recently, McNamara et al conducted a study where software engineers
were explicitly instructed to consider the ACM code of ethics as they were developing
new products. The study found that the code of ethics had no “observed effect” on
their work.16 This suggests that ethical codes of conduct cannot be a solution in and of
themselves, unless accompanied by mechanisms for compliance and accountability.

Lack of transparency in these initiatives is also of foremost concern. While transparency

is a principle that is recommended in multiple ethical frameworks, it ironically
is a feature that most ethics initiatives lack. There is a culture of black box
decision-making in the conversation around AI and ethics, in how ethical boards
are constituted and how principles are developed. When Microsoft constituted
the AI and Ethics in Engineering and Research Committee (AETHER), to make
recommendations on what AI technologies the company should deploy, the
company claimed that “sales” were cut off on the committee’s recommendation.17
Yet the extent, subject, and details of intended sales and nature of deliberation
remain a mystery.

9
This approach also does not account for enforcement or redressal mechanisms,
nor does it contemplate the duty of companies to act in certain ways. Furthermore,
it fails to provide mechanisms by which consumers, public interest advocates, civil
society, and other affected individuals can have agency in the event that companies
fail to meet their own ethical standards.

For instance, following scrutiny over its algorithmic systems,18 Facebook took
several steps towards engaging with ethics initiatives. The company backed an
ethical AI institute at the University of Munich19 and became a founding member of
the Partnership on AI, a multi-stakeholder group that aims to “study and formulate
best practices on AI technologies, to advance the public’s understanding of AI,
and to serve as an open platform for discussion and engagement about AI and its
influences on people and society.”20 Facebook even signed resolutions calling for
the development of ethical principles in the US Congress.21 Yet at the same time,
recent research shows that Facebook discriminates on advertisement delivery on
the basis of gender and race,22 and has also been charged with housing-related
discrimination.23

Facebook’s efforts towards ethical AI had no demonstrable bearing on its practices.

Indeed, these partnerships and other loose commitments did not create any
requirement for the company to follow through on these ideals, nor did they provide
mechanisms to hold the company to account.

As the outcomes of Project Maven show, these initiatives become even more
complex when they are deployed in public-private partnerships, where government
agencies procure privately developed AI systems. In these agreements, sole reliance
on ethical frameworks (as opposed to legally-binding constitutional or human rights
frameworks) dilutes state accountability and rights-based obligations.

As discussed above, there are various cases of governments working together with
industry to improve surveillance of dissidents, precision targeting in drones, or facial
recognition software for law enforcement purposes. These partnerships regularly
take shape in the absence of safeguards or meaningful oversight.

For example, Google’s partnership with the Chinese government to develop a

censored search engine (known as “Project Dragonfly”) would have excluded search
results that were viewed as politically “sensitive” by the Chinese government.24
This represents an egregious violation of international human rights standards
on freedom of expression and information, and also violates Google’s own ethical
principles on AI.

In both public-private initiatives, and corporate-driven efforts, the debate on AI

and ethics is disproportionately influenced by industry initiatives and corporate
aims.25 Even though a variety of actors are developing ethical frameworks, concerns
10
from civil society and academia struggle to get industry support, and even in multi-
stakeholder settings, are easily diluted.

When civil society is invited to partake in deliberation around ethical AI, the division
of seats at the table is not equitable. Ethical initiatives within industry are more often
than not opaque to civil society, with most ethical boards and codes of conduct
being developed and deliberated exclusively in-house. What is more surprising is
that this trend continues even in ethical initiatives convened by governments. For
instance, during deliberations at the European High Level Expert Group on Artificial
Intelligence (EU-HLEG),26 industry was heavily represented, but academics and
civil society did not enjoy the same luxury. And while some non-negotiable ethical
principles were originally articulated in the document, these were omitted from the
final document due to industry pressure.27

National governments have followed a similar pattern. In India, an AI task force

to create a policy and legal framework for the deployment of AI technologies was
constituted without any civil society participation.28 In the United Kingdom, the
Prime Minister’s office for AI has three expert advisors - one professor of computer
science and two industry representatives.29

When we look at efforts led by companies, certain stakeholders appear to use ethics
initiatives as an alternative or preamble to regulation. This approach can carry
dangerous consequences for human rights and the public interest. In proposing
ethical frameworks or principles to avoid regulation under the guise of encouraging
innovation, stakeholders seek to achieve precisely what is discussed above, a
practice Ben Wagner has termed “ethics washing.”30 They affect a veneer of “being
ethical,” yet they have no mechanisms of accountability with which to comply,31 and
thus they face no consequence for their actions. Some may advocate for ethics as a
preamble to regulation, arguing that it is too soon to prescribe regulation addressing
AI.32 But in multiple cases, this has proven to be a strategy of simply buying time to
profit from and experiment on societies and people, in dangerous and irreversible
ways.

Ethics initiatives tend to frame their discussion of AI in terms of companies doing

“the right thing,” in accordance with high level ethical principles or standards. But
we have seen ethical codes of conduct become a smokescreen for doing “the right
thing,” even when there is no clear understanding of what “the right thing” is, or how
to measure it.

For example, Google’s ethical principles laid out its aspirations to build AI systems
that are socially beneficial, and also to avoid creating or reinforcing unfair bias. A
few months later, the company constituted an ethics board, including individuals33
who demonstrably contradicted the basic assumptions behind Google’s ethical
principles. This made the principles and actual company practice fundamentally
11
incongruous, and Google dissolved the board just days later, following public and
internal pushback.34 Google’s AI principles, therefore, have no teeth - they do not
preclude the company from violating its own principles because they create no
obligation or duty in the first place. The link between ethical aspirations and industry
duty is weak at best, and non-existent at worst.

Another problem with trusting companies to do “the right thing” comes from their
lack of understanding of the societal impacts of technology and appropriate
ways to deal with them. For instance, in April 2018, in his testimony before the
United States Congress, Facebook CEO Mark Zuckerberg revealed the company’s
increasing reliance on AI tools to solve problems of hate speech, terrorist
propaganda, election manipulation and misinformation. But research and media
reports have shown that AI tools are ill-suited to do this work - they are not
technically equipped to understand societal nuances or context in speech, and often
make the problem worse.

In fact, prior to the 2017 escalation of military attacks on Rohingya people in

Myanmar, local activists gathered substantial evidence that Facebook was
automatically censoring the word “kalar,” a derogatory local slur used to refer to
Rohingya Muslims. By simply flagging the word as problematic, without accounting
for the immediate context in which it was being used, this step led to the censorship
of numerous posts in which people attempted to discuss use of the term, its history,
and efforts to curb hate speech in the country. Meanwhile, users who wanted to use
the term as an insult simply opted for an alternative spelling. All told, Facebook’s
effort to deploy AI in order to reduce hate speech in this volatile political environment
resulted in censorship of legitimate speech and had no demonstrable effect towards
curbing hate speech.

Finally, on a more granular level, it is also important to note that while ethical
principles are put forth as aspirational goals, even when well-intended and narrowly
tailored, there has seldom been guidance on how to provide balancing mechanisms
for conflicting principles.

2.2 The Technical Approach: Fairness, Accountability and Transparency (FAT)

Widespread use of AI systems on society has brought to the fore concerns around
discrimination,35 injustice,36 the exercise of rights37, amongst others. A community of
researchers, academics, and scientists have been working to address these issues
by developing AI systems that are fair, accountable, and transparent (FAT). This
particular field of technical work has grown over decades, and pre-dates the current
resurgence in AI-focused ethics initiatives.

12
The notion of fairness in machine learning (the most popular subset of AI
techniques) is arguably the most prominent topic in the field today, with researchers
and practitioners attempting to articulate what fairness entails and in turn,
operationalise these learnings at the time of deployment.

Fairness is a normative concept and can be defined in multiple ways.38 Fairness

pushes technical thinking beyond just optimisation, to consider the actual impact of
AI systems and their implications for society. But this is not without complication.
Sometimes one type of fairness can be more appropriate than the other,
necessitating a tradeoff. This balancing act requires some guidance. There is also
a question of tradeoffs between fairness and other values - a perfectly fair system
may not be very accurate, and accuracy can sometimes undermine fairness.39

The ability of AI systems to be at once invisible, opaque, and inscrutable

also gave rise to efforts to make them accountable and transparent. This is
especially important given the increasing use of AI systems in the public sector
for consequential decision-making. The overarching assumption here is that
transparency affords enough insight into a system, and in turn enables explanation,
redressal, appeal, and accountability to follow. The precise feasibility of this
approach is the subject of ongoing debate.40 Some experts believe that peering
inside black boxes could lend some insight into the inner functioning of a system,
which could then lead to the ability to hold them to account.41 Others believe that
given the evolving nature of AI systems, transparency on its own will not help.42

Calls for accountability, however, go beyond this assumption. AI systems

are increasingly used to carry out consequential decision making and deliver
services that were once the responsibility of states. Accountability is essential in
relationships with power differentials, making this question both technically and
legally crucial.

The conversation around FAT has been led by academia for many years, and this is
still true today. The Association for Computing Machinery Fairness, Accountability,
and Transparency (ACM FAT*) in ML43 is perhaps the most popular venue for FAT
work in the field. Some recent papers from the event included work on improving
fairness of facial recognition algorithms,44 distinguishing between fairness and bias,45
and a study on bias in news and fact checking. 46

Industry has also engaged with the idea of fairness. In 2018, Accenture published
a fairness toolkit to help businesses work towards fair outcomes in the process of
deploying AI systems.47 Spotify and Microsoft researchers recently presented work
on the challenges they face when trying to implement fairness on a daily basis for
technical experts working on FAT issues.48 FAT work is also carried out in industry
consortiums like the PAI and IEEE, which often pave the way for wider stakeholder
engagement with these issues. Venues such as the PAI offer dedicated working
13
groups on FAT issues, and have strong civil society participation. Other venues like
ACM FAT* are still very much technical, leaning strongly towards academia.

2.2.3 Critical gaps in the FAT approach

The design, development, deployment, and assessment of AI systems can be

profoundly affected by work towards fairness, accountability, and transparency.
While this was traditionally a purely technical field of research, there is growing
acknowledgement of the need to bring in expertise from disciplines beyond
computer science and mathematics. Experts from law, social science, philosophy,
and other disciplines are beginning to engage with the field, and to inform and
enhance these concepts.

Even as this field pushes the limits of our current understanding of fair, accountable
and transparent AI systems, ARTICLE 19 believes that some concerns with the
current approach remain.

First, technical work needs to imbibe socio-political awareness in a way that

is both meaningful and competent. Many AI systems currently in use make
assumptions about social, cultural, and political values that reflect a lack of worldly
expertise. Optimising for a particular type of output is not enough if the surrounding
societal and historical context is not brought to bear in tandem with technical
considerations. No matter how precise a particular technical definition may be,
an accurate technical output does not necessarily generate a fair socio-technical
outcome for the user or people affected by the technology.49

Consider how a credit scoring model might be trained to operate. How could such
a model affect historically disadvantaged populations? It is possible that, when
studied in isolation, historical discrimination of the past could easily translate
to future discrimination. Recognition of structural and systemic inequalities,
constitutional guarantees of affirmative action (where applicable), and the
responsibility to correct past practices would mean building a technology that
transcends this view. This has been recognised by the community itself and is
slowly being addressed. Some of the most recent literature in this field attempts to
learn and further develop these concepts by borrowing from political philosophy,50
social sciences,51 and even the fields of education and hiring.52

But significant questions remain: Which values are embedded? How are technical
experts positioned to understand them? What shared terms exist around these
values, and what are possible ways to codify them?

Second, the FAT approach does not structurally engage with responsibility, rights,
duties, or harms of technical systems in an actionable way. This it shares in
common with the ethics initiatives discussed in the previous section. The FAT

14
approach does not necessarily articulate responsibility, harm, or expectation of fair
treatment to ensure that these goals are met.

For example, in building a predictive policing algorithm, beyond articulating ways in

which it is accountable, and the constraints taken into consideration to ensure it will
not be unfair or discriminatory, the development of these systems should engage
with certain key questions: Who can demand accountability of these systems? Who
are the people to whom fairness and transparency is owed? Which authorities and
institutions will be accountable?

Also similar to the ethics approach, the FAT approach does not address the
tangible effect of technical systems on the exercise of rights, or on people’s lives.
For instance, there has been recent pressure on technology companies like IBM
and Amazon (among others) to ensure that facial recognition systems have equal
accuracy rates between vulnerable and dominant racial groups.53 Ensuring this could
satisfy some definitions of fairness, yet it does not engage with the broader question
in play. Are facial recognition systems a threat to the exercise of fundamental rights
such as privacy and free expression? Does a perfectly fair facial recognition system
-- one that is equally and similarly accurate across demographic groups -- have a
disproportionate impact on vulnerable groups that have historically been subject to
surveillance? Should the faces of individuals from these groups be used in training
datasets? What implications does this have on their autonomy and privacy?

Relatedly, the FAT approach has been built to ensure that systems are fair,
transparent and accountable, but it does not empower individuals or surrounding
institutional mechanisms to challenge the decisions that these systems make. By
sidestepping the question of rights and duties, affected individuals and oversight
authorities cannot meaningfully challenge systems that fail them. Take for instance
New York’s Automated Decision Systems Task Force that examines the use of
automated decision-making systems by the city to prevent bias or other harms.
The task force has adopted principles of equity, fairness and accountability,54 and
has leading AI scientists and researchers among its members. Yet a full year after it
was constituted, the task force is unable to carry out its duties because there are no
accompanying institutional mechanisms, oversight powers, or rights to investigate
uses of automated decision-making systems.55

15
3. Towards a human rights-based
approach
Alongside technical and ethical approaches to addressing the societal impact of AI
systems, there has been a third approach, focused on law and regulation. While a
handful of states have begun to explore legal frameworks for governing AI, various
stakeholders have turned to international human rights instruments as a way
forward.

There have been preliminary attempts to begin regulating AI at national and regional
levels. In the EU, for example, the General Data Protection Regulation (GDPR)56
articulates a few rights with respect to automated decision making. Under the GDPR,
data controllers (typically companies or state entities) are required to provide data
subjects with information about “the existence of automated decision making,
including profiling...and, at least in those cases, meaningful information about the
logic involved, as well as the significance and the envisaged consequences of such
processing for the data subject.”57 The GDPR also provides that individuals have
the “right not to be subject to a decision based solely on automated processing,
including profiling, which produces legal effects concerning him or her or similarly
significantly affects him or her.”58 and guarantees that data subjects can seek
human intervention and contest the decision.59 The extent and impact of these
provisions are the subject of ongoing debate, and are sometimes referred to as
the “right to explanation.”60 In the United States, an Algorithmic Accountability
Bill introduced in the Senate in April 2019 contemplates impact assessments for
automated decision systems and data protection.61

Given current isolated regional and national attempts at regulation of AI systems,

the relevance and importance of internationally recognised, legally binding human
rights at the national and international levels cannot be overstated. There has been
substantial work on the intersection of AI and human rights so far, particularly
around freedom of expression, privacy,62 non-discrimination, and governance.63 The
Toronto Declaration on protecting the right to equality and non-discrimination in
machine learning systems,64 launched in May 2018, aims “to underline the centrality
of the universal, binding and actionable body of human rights law and standards,
which protect rights and provide a well-developed framework for remedies.”

More recently, over 100 organisations signed a statement focused on civil liberties
concerns regarding the use of pretrial risk assessment tools.65 The United Nations
also has weighed in on the human rights debate: UN Special Rapporteur on
the promotion and protection of the right to freedom of opinion and expression
presented a detailed report on the human rights impact of AI systems to the General
Assembly.66
16
Some national governments also have adopted a human rights-based approach. In
July 2018, the Australian human rights commission launched a three-year project
to understand the human rights impact of AI.67 The governments of Canada and
France are steering an international study group aimed at human centric artificial
intelligence, using human rights as one of the anchors of the study.68 There has also
been work highlighting the human rights obligations of businesses in the context of
AI.69

Having discussed the current deliberation around societal impacts of AI, we now
wish to propose a way forward. ARTICLE 19 believes that current deficiencies
discussed above -- the lack of enforcement, accountability, meaningful redressal,
and individual empowerment -- could be constructively supplanted through a
human rights-based approach.

The human rights-based approach identifies rights holders (people who use or are
affected by technologies) and duty bearers (companies or governments deploying
said technologies). It is a universal set of principles, has binding effect, and is based
on the rule of law. It draws on an internationally recognised system of law that
defines both business and state responsibility and the specific standards they must
adhere to. This means that rights, reasonable restrictions, their status under law
and implementation in practice, are anchored in a system that is verifiable, specific,
and detailed. This international system has grounding in law, is based on commonly
understood language and affords procedures and institutions that can help ensure
that duty bearers meet their obligations, and that rights holders have recourse to
effective remedies.70

What makes a human rights-based approach particularly important in the context

of AI is the pre-existing focus on the use of privately developed technology by
states. The nexus between states and businesses is explicitly addressed in the
United Nations Guiding Principles on Business and Human Rights (UNGPs).71 These
safeguards recognise the possibility of states shirking their obligations under
international law by contracting with private actors, and they provide mechanisms to
ensure the protection and fulfillment of human rights obligations. It is crucial to note
that human rights are both a legal and ethical standard.

In this section, we will contemplate how a human rights-based approach can benefit
existing technical and ethical conversations discussed above, and in turn also
identify ways in which the converse is also true, i.e. what aspects of the FAT and
ethics conversations can inform existing human rights approaches to AI.

17
3.1 A human rights-based approach to Ethics

In our analysis, our primary concern with ethics initiatives was their common lack of
accountability and enforcement mechanisms. Some appeared to be little more than
efforts to skirt regulation, while others, though perhaps well-intended, did not have
proverbial teeth: They put forth admirable goals, but offered little if any mechanism
of accountability or enforcement. If a human rights-based approach were brought to
bear here, it could complement ethical principles with an enforcement mechanism,
drawing from the UNGPs, which offer comprehensive guidance on how to make
industry practices more concretely accountable, in addition to mechanisms for
redressal, safeguarding rights, and ensuring performance of corresponding duties.

Grounding ethical principles in human rights standards would also preclude the
worrying prospect of “ethics washing” and rubber stamping. Human rights law has a
legacy of constitutional and judicial interpretation, legal oversight, and enforcement
mechanisms that are subject to review.

We also observed that industry initiatives and corporate aims have come to
dominate and set the agenda within many ethics initiatives. A human rights-
based approach could re-balance the scales, as it would embrace the realities of
carrying out business by prescribing a specific, detailed account of what rights and
obligations are in play. It would also keep individuals and the rights owed to them as
the central focus, and the point for calibration.

Perhaps as a result of the agenda-setting by industry, the ethics framework does not
contemplate the duty of companies to act in certain ways. Instead, it tends to focus
on responsibility for outcomes (who holds it, and what it entails), as opposed to
addressing bigger questions around business models and the monopolistic power of
tech companies. A human rights-based approach would correct for this by targeting
every step of a company’s business model, thus bringing into focus not just outputs
from a particular business, but also processes and safeguards for rights holders and
duty bearers. This could also be drawn from the UNGPs, alongside international and
national law, if applicable.

Finally, we touched upon the the risk (common in multi-stakeholder initiatives) of

developing recommendations that are internally contradictory and do not provide
balancing mechanisms for conflicting principles. A rights-based approach would
provide guidance on how to balance principles as it would be informed by the
reasonable restrictions on particular rights, the contours within which the rights
stand, and how these conflicts have been treated in the law. Finding parallels in law
can be immensely helpful.

18
3.2 A human rights-based approach to FAT

In the section discussing FAT above, we identified a few key issues where this work
could be further developed to incorporate the social, political, and legal context in
which AI systems are deployed.

A human rights-based approach could strongly complement the technical specificity

of the FAT approach by providing a robust legal framework that is grounded
in the relationship between rights holders and duty bearers, and articulates
standards for enforcement, accountability, meaningful redressal, and individual
empowerment. Technical definitions of fairness, transparency and accountability
tend to focus mostly on outcomes in a given situation, not on the process of
reaching the outcome. A rights-based approach would consider due process and
non-discrimination in ways that transcend just outputs. ARTICLE 19 believes that
a human rights-respecting AI system is not simply one that does not produce
rights-violating outputs, but that it is human rights-respecting through design,
development and deployment.

Invoking a human rights-based approach would allow FAT proponents to make

a shift from precise technical models to fair sociotechnical systems. Human
rights-based approaches encourage an understanding of context, rights, and
corresponding duties in a manner that necessitates taking into account social
realities and institutions. The human rights framework also provides a baseline
acknowledgement of the scope and application of each right.

In turn, this could allow the FAT field to more meaningfully address the tangible
effects of AI systems on the people’s lives and fundamental rights. Because human
rights-based approaches are grounded in the relationship between rights holders
and duty bearers, it refocuses these questions in terms of individual and collective
rights, and the obligations owed to rights holders, including rights to seek redressal.

The FAT approach does not empower individuals, or provide particular agency in
this regard. A human rights-based approach offers enforcement mechanisms, and
prescribes institutional processes to work with in the event of rights violations,
in addition to offering existing external mechanisms. In case of discrimination,
for instance, the FAT approach may lay out tests for fairness and methods of
accountability, and a rights-based framework augments these by also providing
redressal mechanisms for people affected.

The FAT approach focuses on systems, and mechanisms for accountability around
systems, but does not clearly articulate responsibility, harm, and expectation of fair
treatment. A human rights-based approach is particularly well positioned to fill this
gap as it is informed by international legal standards and clear articulation of roles
and responsibilities.
19
A human rights-based approach could also help to carry forward questions in the
FAT field about how to engage with and critically examine values. Amid efforts
to promote value-sensitive design72 of machine learning systems, and a growing
awareness of the need to explicitly deal with values that are encoded in systems,
human rights provide a universal set of values with legal grounding. ARTICLE 19
believes that invoking a human rights framework is especially important given these
learnings, as human rights are the most universal set of values that we have, with
shared language and decades of interpretation and implementation.

Finally, a human rights-based approach would provide guidance on how to balance

rights and values, or how to navigate tradeoffs between say, fairness and accuracy.
Under a human rights-based approach, these would be informed by reasonable
restrictions on particular rights, and how similar scenarios have been treated in case
law.

3.3 How might FAT and ethics initiatives help improve a human rights-based
approach?

Advocacy for human rights protections should take into account technical
considerations: While human rights have grounding in law, and also universally
understood language and meaning, advocates for human rights could strengthen
their approach to AI systems by learning and taking into account technical
necessities, limitations, and terminologies. This would not only enable precision
across disciplines to emerge, it would also carry out important translation between
technical and non-technical audiences.

A deeper understanding of technical considerations could help as well. There

are some inevitable tradeoffs that must be made in the process of developing AI
systems. For instance, the tradeoff between fairness and accuracy is a common
dilemma. A human rights-based approach would be strengthened by understanding
the considerations that go into such tradeoffs.

Finally, while human rights serve as a minimum requirement for AI systems to

adhere to, a human rights-based approach could be strengthened by also taking into
consideration aspirational goals that go beyond just this minimum requirement, as
ethical principles do.

20
4. Recommendations
We call on industry to:

1. Affirm their commitment to the UN Guiding Principles on business and

human rights.

2. Embed international human rights standards, particularly those of protecting

freedom of expression and privacy, in the development and deployment of AI
systems in a manner that lays out precise obligations and commitments for
companies building these technologies.

3. Explicitly invoke a human rights-based approach to ethical codes of conduct;

identify duty bearers and rights holders.

4. Publicly articulate what safeguards are in place to monitor and evaluate

ongoing ethical efforts, and institute accountability and redressal
mechanisms.

5. Enhance transparency around ethical AI initiatives, from the constitution

of ethical boards (including their powers and functions), to how ethical
principles are internally evaluated, to the impact of these initiatives on
businesses.

6. Ensure that AI systems undergo rigorous human rights impact assessments,

and create feedback and continuous auditing mechanisms for the same.

7. As a continuation of the above, meaningfully engage with civil society and

academia at each stage of the process, to cultivate constructive criticism as
part of internal deliberation processes.

We call on states to:

1. Reaffirm their commitment to international human rights standards,

particularly those protecting freedom of expression and privacy, and let this
commitment guide their specific national AI strategies.

2. Ensure that AI systems in the public sector undergo adequate human rights
impact assessments, due diligence, and continuous auditing. These are not
systems that can simply be rolled out. They should instead be tailored to the
exact context and use for which they are intended.

21
 3. Root the design, development, and deployment of AI systems in
constitutional guarantees and human rights standards.

4. Hold AI systems to accountability, responsibility, and constitutional

standards without dilution or exception.

5. Ensure that national and international efforts around AI are equally informed
by human rights concerns, constitutional standards, and the public interest
as they are by industry concerns.

6. Establish necessary infrastructure and resources to provide remedy for

human rights violations caused by AI systems.

We call on civil society to:

1. Work towards greater transparency and accountability of mechanisms

(such impact assessments, the Universal Periodic Review, and rights-based
mechanisms at national levels) and institutions (like the Human Rights
Council, General Assembly, Special Rapporteurs, national and regional
institutions that focus on compliance with and enforcement of human rights)
that make up a rights-based approach.

2. Advocate for the promotion and protection of human rights in the context of
AI systems in a manner that is also informed by technical considerations
and limitations.

3. Actively engage with ethical initiatives to integrate strong human rights

language and protections into their substantive and procedural functioning.

22
Endnotes
1 In a previous report co-written with 5 Barocas, Solon, and Andrew D. Selbst.
Privacy International, we attempted ‘Big Data’s Disparate Impact’. California
to capture this new technology in Law Review 671 (2016). https://papers.
the following definition: “The term ssrn.com/sol3/Papers.cfm?abstract_
‘AI’ is used to refer to a diverse range id=2477899.
of applications and techniques, at
different levels of complexity, autonomy 6 Veale Michael, Binns Reuben, and
and abstraction. This broad usage Edwards Lilian. ‘Algorithms That
encompasses machine learning (which Remember: Model Inversion Attacks
makes inferences, predictions and and Data Protection Law’. Philosophical
decisions about individuals), domain- Transactions of the Royal Society A:
specific AI algorithms, fully autonomous Mathematical, Physical and Engineering
and connected objects and even the Sciences 376, no. 2133 (November
futuristic idea of an AI ‘singularity’.” We 2018). https://doi.org/10.1098/
also outlined a number of key concepts rsta.2018.0083.
related to AI systems. ARTICLE 19,
and Privacy International. ‘Privacy and 7 Cowls, Josh, and Luciano Floridi.
Freedom of Expression In the Age of ‘Prolegomena to a White Paper on an
Artificial Intelligence’, 2018. https:// Ethical Framework for a Good AI Society’,
www.article19.org/wp-content/ (June 2018). https://papers.ssrn.com/
uploads/2018/04/Privacy-and- abstract=3198732.
Freedom-of-Expression-In-the-Age-of-
Artificial-Intelligence-1.pdf. 8 Floridi, et al. ‘An Ethical Framework for
a Good AI Society: Opportunities, Risks,
2 Vincent, James. ‘Artificial Intelligence Principles, and Recommendations’.
Is Going to Supercharge Surveillance’. In Minds and Machines (December
The Verge, January 2018. https://www. 2018). https://www.researchgate.net/
theverge.com/2018/1/23/16907238/ publication/328699738_An_Ethical_
artificial-intelligence-surveillance- Framework_for_a_Good_AI_Society_
cameras-security. Opportunities_Risks_Principles_and_
Recommendations.
3 Coughlan, Sean. ‘Google “to End
Pentagon AI Project”’, June 2018, sec. 9 Dutton, Tim. ‘An Overview of National
Business. https://www.bbc.com/news/ AI Strategies’. Politics + AI (blog), June
business-44341490. 2018. https://medium.com/politics-ai/
an-overview-of-national-ai-strategies-
4 Kania, Elsa B. ‘China’s AI Giants Can’t Say 2a70ec6edfd.
No to the Party’. Foreign Policy (blog), August
2018. https://foreignpolicy.com/2018/08/02/
chinas-ai-giants-cant-say-no-to-the-party/.
23
10 ‘Ethics guidelines for Trustworthy April 2019. https://people.engr.ncsu.edu/
AI’. Accessed 15 April 2019. https:// ermurph3/papers/fse18nier.pdf.
ec.europa.eu/digital-single-market/en/
news/ethics-guidelines-trustworthy-ai. 17 Boyle, Alan. ‘AI Expert Says Microsoft
Is Cutting off Some Sales Due to Ethics
11 ‘Asilomar AI Principles’. Future of Concerns – GeekWire’. Geekwire, April
Life Institute. Accessed 11 April 2019. 2018. https://www.geekwire.com/2018/
https://futureoflife.org/ai-principles/. microsoft-cutting-off-sales-ai-ethics-
top-researcher-eric-horvitz-says/.
12 ARTICLE 19. ‘Submission of
Evidence to the House of Lords Select 18 Madrigal, Alexis C. ‘What Facebook
Committee on Artificial Intelligence’, Did to American Democracy’. The
September 2017. https://www.article19. Atlantic, October 2017. https://
org/wp-content/uploads/2017/10/ www.theatlantic.com/technology/
ARTICLE-19-Evidence-to-the-House- archive/2017/10/what-facebook-
of-Lords-Select-Committee-AI-1. did/542502/. Also see: Angwin, Julia,
pdf. Also see ‘High-Level Expert Group Madeleine Varner and Ariana Tobin. ‘
on Artificial Intelligence | Digital Single Facebook Enabled Advertisers to Reach
Market’. Accessed 11 April 2019. https:// ‘Jew Haters’’. ProPublica, September
ec.europa.eu/digital-single-market/ 2017. https://www.propublica.org/
en/high-level-expert-group-artificial- article/facebook-enabled-advertisers-
intelligence. to-reach-jew-haters.

13 ‘Partners - The Partnership on AI’. 19 ‘Facebook and the Technical

Accessed 11 April 2019. https://www. University of Munich Announce New
partnershiponai.org/partners/. Independent TUM Institute for Ethics
in Artificial Intelligence | Facebook
14 ARTICLE 19. ‘Google: New Guiding Newsroom’. Accessed 11 April
Principles on AI show progress but still 2019. https://newsroom.fb.com/
fall short on human rights protections’, news/2019/01/tum-institute-for-ethics-
June 2018. https://www.article19. in-ai/.
org/resources/google-new-guiding-
principles-on-ai-show-progress- 20 ‘About - The Partnership on AI’.
but-still-fall-short-on-human-rights- Accessed 11 April 2019. https://www.
protections/. partnershiponai.org/about/.

15 Google. ‘Our Principles’| Google 21 ‘Ro Khanna and Brenda l. Lawrence

AI. Accessed 11 April 2019. https:// Introduce Resolution Calling for the Ethical
ai.google/principles/. Development of Artificial Intelligence’.
Congressman Ro Khanna, February 2019.
16 McNamara, Andrew, Justin Smith and https://khanna.house.gov/media/press-
Emerson Murphy Hill. ‘Does ACM’s Code releases/release-ro-khanna-and-brenda-
of Ethics Change Ethical Decision Making l-lawrence-introduce-resolution-calling-
in Software Development?. Accessed 15 ethical.
24
22 Ali, Muhammad, Piotr Sapiezynski, 27 Dialogue Seminar on Artificial
Miranda Bogen, Aleksandra Korolova, Intelligence: Ethical Concerns -
Alan Mislove, and Aaron Rieke. Comments by T. Metzinger, 2019. http://
‘Discrimination through Optimization: www.europarl.europa.eu/streaming.
How Facebook’s Ad Delivery Can Lead
to Skewed Outcomes’. ArXiv:1904.02095 28 ‘Members AI Task Force (AITF)’.
[Cs], (April 2019. http://arxiv.org/ Accessed 11 April 2019. https://www.
abs/1904.02095. aitf.org.in/members.

23 ‘HUD versus Facebook Charge’, 29 Williams, Oscar. ‘DeepMind’s

August 2018. https://www.hud.gov/ Demis Hassabis Is Set to Advise the
sites/dfiles/Main/documents/HUD_v_ Government’s New Office for AI’. NS
Facebook.pdf. Please note, Facebook Tech (blog), June 2018. https://tech.
recently settled these challenges. newstatesman.com/business/demis-
Please see: Levy, Pema. ‘Facebook hassabis-office-ai-adviser.
Settles Civil Rights Lawsuits Over Ad
Discrimination’. Mother Jones, March 30 Wagner, Ben. ‘Ethics as an Escape
2019. https://www.motherjones. from Regulation: From ethics-washing
com/politics/2019/03/facebook- to ethics-shopping?’ . M. Hildebrandt
settles-civil-rights-lawsuits-over-ad- (Ed.), Being Profiling. Cogitas ergo sum.
discrimination/. Amsterdam University Press (2018).
https://www.privacylab.at/wp-content/
24 Gallagher, Ryan. ‘Google Plans to uploads/2018/07/Ben_Wagner_Ethics-as-
Launch Censored Search Engine in an-Escape-from-Regulation_2018_BW9.pdf.
China, Leaked Documents Reveal’. The
Intercept (blog), August 2018. https:// 31 Nemitz, Paul. ‘Constitutional
theintercept.com/2018/08/01/google- Democracy and technology in the age
china-search-engine-censorship/. of artificial intelligence’. Philosophical
Transactions of the Royal Society A:
25 Cath, Corinne. ‘Who Is Driving the Mathematical, Physical and Engineering
AI Agenda and What Do They Stand Sciences 376, no. 2133 (November
to Gain?’ NS Tech (blog), July 2018. 2018). https://doi.org/10.1098/
https://tech.newstatesman.com/ rsta.2018.0089.
guest-opinion/regulating-artificial-
intelligence-ai. 32 Kharpal, Arjun. ‘Intel CEO Brian
Krzanich: It’s Too Early to Regulate AI’,
26 ‘High-Level Expert Group on Artificial November 2017. https://www.cnbc.
Intelligence | Digital Single Market’. com/2017/11/07/ai-infancy-and-too-
Accessed 11 April 2019. https:// early-to-regulate-intel-ceo-brian-
ec.europa.eu/digital-single-market/ krzanich-says.html.
en/high-level-expert-group-artificial-
intelligence.

25
33 Smith, Reiss. ‘Google appoints 39 See, for instance, Kleinberg, Jon,
transphobic conservation to AI ethics Sendhil Mullainathan, and Manish
board’. Pink News, March 2019. https:// Raghavan. ‘Inherent Trade-Offs in the
www.pinknews.co.uk/2019/03/27/ Fair Determination of Risk Scores’.
google-ai-artificial-intelligence-ethics- ArXiv:1609.05807 [Cs, Stat], (September
board-kay-coles-james/. 2016). http://arxiv.org/abs/1609.05807.

34 Piper, Kelsey. ‘Google cancels ethics 40 Marda, Vidushi. ‘Machine Learning

board in response to outcry’. Vox, April and Transparency: A Scoping Exercise’,
2019. https://www.vox.com/future- (November 2017. https://papers.ssrn.
perfect/2019/4/4/18295933/google- com/abstract=3236837.
cancels-ai-ethics-board.
41 Diakopoulos, Nick and
35 Weissmann, Jordan. ‘Amazon Created Michael Koliska. ‘Algorithmic
a Hiring Tool Using AI. It Immediately Transparency in the News Media’.
Started Discriminating Against Women.’ In Digital Journalism (2016), DOI:
Slate Magazine, October 2018. https:// 10.1080/21670811.2016.1208053.
slate.com/business/2018/10/
amazon-artificial-intelligence-hiring- 42 Kroll, Joshua A, et al. ‘Accountable
discrimination-women.html. Algorithms’. In University of
Pennsylvania Law Review, Vol. 165, 2017
36 Angwin, Julia et. al. ‘Machine Bias’. Forthcoming. https://papers.ssrn.com/
ProPublica, May 2016. https://www. sol3/papers.cfm?abstract_id=2765268.
propublica.org/article/machine-
bias-risk-assessments-in-criminal- 43 ACM Conference on Fairness,
sentencing. Accountability, and Transparency
(FAT*). Accessed 15 April, 2019. https://
37 ARTICLE 19, and Privacy International. fatconference.org/.
‘Privacy and Freedom of Expression In
the Age of Artificial Intelligence’, 2018. 44 Buolamwini, Joy, and Timnit
https://www.article19.org/wp-content/ Gebru. ‘Gender Shades: Intersectional
uploads/2018/04/Privacy-and- Accuracy Disparities in Commercial
Freedom-of-Expression-In-the-Age-of- Gender Classification’. In Conference
Artificial-Intelligence-1.pdf. on Fairness, Accountability and
Transparency, 2018, 77–91. http://
38 Narayanan, Aravind. ‘FAT* Tutorial: proceedings.mlr.press/v81/
21 Fairness Definitions and Their buolamwini18a.html.
Politics’. Google Docs, February
2018. https://docs.google.com/ 45 Metcalf, Jacob. Engineering for
document/d/1bnQKzFAzCTcBcNvW5 Fairness: How a Firm Conceptual
tsPuSDje8WWWY-SSF4wQm6TLvQ/ Distinction between Unfairness and Bias
edit?usp=sharing&usp=embed_ Makes it Easier to Address Un/Fairness.
facebook. Conference on Fairness, Accountability,
and. Transparency, New York,
26
Forthcoming Proceedings of Machine Accountability, and. Transparency,
Learning Research (March 2018). New York, Forthcoming Proceedings
http://ethicalresolve.com/content/ of Machine Learning Research Vol. 81
uploads/2019/01/fat2019tutorials- (December 2017): 1–11. https://papers.
paper26.pdf. ssrn.com/sol3/papers.cfm?abstract_
id=3086546.
46 Babaei, Mahmoudreza et al,
‘Analyzing Biases in Perception of 51 Miller, Tim. ‘Explanation in Artificial
Truth in News Stories and Their Intelligence: Insights from the Social
Implications for Fact Checking’. Sciences’. ArXiv:1706.07269 [Cs], (June
In FAT* ‘19 Proceedings of the 2017. http://arxiv.org/abs/1706.07269.
Conference on Fairness, Accountability,
and Transparency Pages 139-139 52 Hutchinson, Ben, and Margaret
(2019). https://dl.acm.org/citation. Mitchell. ‘50 Years of Test (Un)
cfm?id=3287581. Fairness: Lessons for Machine
Learning’. Proceedings of
47 Lomas, Natasha. ‘Accenture Wants the Conference on Fairness,
to Beat Unfair AI with a Professional Accountability, and Transparency -
Toolkit’. TechCrunch, 2018. http:// FAT* ’19, (2019, 49–58. https://doi.
social.techcrunch.com/2018/06/09/ org/10.1145/3287560.3287600.
accenture-wants-to-beat-unfair-ai-
with-a-professional-toolkit/. 53 Buolamwini, Joy, and Timnit
Gebru. ‘Gender Shades: Intersectional
48 Cramer, Henriette, et al. ‘Challenges Accuracy Disparities in Commercial
of incorporating algorithmic ‘fairness’ Gender Classification’. In Conference
into practice’. In Conference on Fairness on Fairness, Accountability and
Accountability, and Transparency Transparency, 2018, 77–91. http://
Tutorial (February 2019). https://drive. proceedings.mlr.press/v81/
google.com/file/d/1rUQkVS0NzSH3IEq buolamwini18a.html.
ZDsczSxBbhYHbjamN/view.
54 ‘Mayor de Blasio Announces First-
49 Marda, Vidushi. ‘Artificial Intelligence In-Nation Task Force To Examine
Policy in India: A Framework for Automated Decision Systems Used By
Engaging the Limits of Data-Driven The City’. The Official Website of the
Decision-Making’. Philosophical City of New York, May 2018. https://
Transactions of the Royal Society A: www1.nyc.gov/office-of-the-mayor/
Mathematical, Physical and Engineering news/251-18/mayor-de-blasio-first-in-
Sciences 376, no. 2133 (November nation-task-force-examine-automated-
2018). https://doi.org/10.1098/ decision-systems-used-by.
rsta.2018.0087.
55 Stoyanovich, Julia, and Solon
50 Binns, Reuben. ‘Fairness in Machine Barocas. Testimony of Julia Stoyanovich
Learning: Lessons from Political and Solon Barocas before New York
Philosophy’. Cnference on Fairness, City Council Committee on Technology,
27
regarding Update on Local Law 49 https://papers.ssrn.com/sol3/papers.
of 2018 in Relation to Automated cfm?abstract_id=3039125.
Decision Systems (ADS) Used by
Agencies, § New York City Council 61 Algorithmic Accountability Act, 2019.
Committee on Technology (/2019). Accessed April 15, 2019. https://www.
https://dataresponsibly.github.io/ wyden.senate.gov/imo/media/doc/
documents/StoyanovichBarocas_ Algorithmic%20Accountability%20
April4,2019testimony.pdf. Act%20of%202019%20Bill%20Text.pdf.

56 Regulation (EU) 2016/679 62 ARTICLE 19, and Privacy International.

General Data Protection ‘Privacy and Freedom of Expression In
Regulation (GDPR), https://eur- the Age of Artificial Intelligence’, 2018.
lex.europa.eu/legal-content/en/ https://www.article19.org/wp-content/
TXT/?uri=CELEX%3A32016R0679. uploads/2018/04/Privacy-and-
Freedom-of-Expression-In-the-Age-of-
57 Regulation (EU) 2016/679, Article Artificial-Intelligence-1.pdf.
13(2)(f), 14(2)(g), 15(1)(h).
63 Latonero, Mark. ‘Governing Artificial
58 Regulation (EU) 2016/679, Article Intelligence - Upholding Human Rights
22(1). and Dignity’. Data & Society, October
2018. https://datasociety.net/output/
59 Regulation (EU) 2016/679, Article governing-artificial-intelligence/.
22(1) - 22(4).
64 ‘The Toronto Declaration: Protecting
60 Wachter, Sandra, Brent Mittelstadt the Rights to Equality and Non-
and Luciano Floridi. ‘Why a Right to Discrimination in Machine Learning
Explanation of Automated Decision- Systems - Access Now’. Accessed 11
Making Does Not Exist in the April 2019. https://www.accessnow.org/
General Data Protection Regulation’ the-toronto-declaration-protecting-
In International Data Privacy Law the-rights-to-equality-and-non-
(December 2017). https://papers. discrimination-in-machine-learning-
ssrn.com/sol3/papers.cfm?abstract_ systems/.
id=2903469. Also see Edward,
Lilian, and Michael Veale. ‘Slave to 65 ‘Pretrial Risk Assessments’. The
the Algorithm? Why a ‘Right to an Leadership Conference Education
Explanation’ Is Probably Not the Remedy Fund. Accessed 11 April 2019. https://
You Are Looking For’. In 16 Duke civilrights.org/edfund/pretrial-risk-
Law & Technology Review 18 (2017). assessments/.
https://papers.ssrn.com/sol3/papers.
cfm?abstract_id=2972855. Also see 66 Kaye, David. ‘OHCHR | Report of
Selbst, Andrew D. and Julia Powles. the Special Rapporteur to the General
‘Meaningful Information and the Right Assembly on AI and Its Impact on
to Explanation’. In International Data Freedom of Opinion and Expression’
Privacy Law, vol. 7(4), 233-242 (2017). (United Nations Office of the High
28
Commissioner of Human Rights, 2018). 72 Dobbe, Roel, and Morgan Ames.
https://www.ohchr.org/EN/Issues/ ‘Up Next For FAT*: From Ethical
FreedomOpinion/Pages/ReportGA73.aspx. Values To Ethical Practices’. Medium,
February 2019. https://medium.com/@
67 Australian Human Rights roeldobbe/up-next-for-fat-from-
Commission. ‘Protecting Human Rights ethical-values-to-ethical-practices-
in the Era of Artificial Intelligence’, July ebbed9f6adee.
2018. https://www.humanrights.gov.au/
news/stories/protecting-human-rights-
era-artificial-intelligence.

68 Simonite, Tom. ‘Canada, France Plan

Global Panel to Study the Effects of AI’.
Wired, December 2018. https://www.
wired.com/story/canada-france-plan-
global-panel-study-ai/.

69 Allison-Hope, Dunstan and Hodge,

Mark. ‘Artificial Intelligence: A Rights-
Based Blueprint for Business’. Business
for Social Responsibility BSR, August
2018. https://www.bsr.org/reports/
BSR-Artificial-Intelligence-A-Rights-
Based-Blueprint-for-Business-
Paper-01.pdf.

70 Veen, Christiaan van, and Corinne

Cath. ‘Artificial Intelligence: What’s
Human Rights Got To Do With It?’ Data
& Society: Points, May 2018. https://
points.datasociety.net/artificial-
intelligence-whats-human-rights-got-
to-do-with-it-4622ec1566d5.

71 United Nations Human Rights Office

of the Human Rights Commissioner.
‘Guiding Principles on Business and
Human Rights’, 2011. https://www.
ohchr.org/Documents/Publications/
GuidingPrinciplesBusinessHR_EN.pdf.

29
ARTICLE 19 Free Word Centre 60 Farringdon Road London EC1R 3GA
T +44 20 7324 2500 F +44 20 7490 0566
E info@article19.org W www.article19.org Tw @article19org facebook.com/article19org