Unite Paper 2021(1)

A Framework for Ethical AI at the United Nations

Prepared by: Lambert Hogenhout

Organization: UN Office for Information and Communications Technology
Contact: [email protected]
Date: 15-3-2021

Contents

EXECUTIVE SUMMARY 2
INTRODUCTION 3
1. PROBLEMS WITH AI 5
2. DEFINING ETHICAL AI 11
3. IMPLEMENTING ETHICAL AI 20
CONCLUSION 23

Unite Papers— “Informing and Capturing UN Technological Innovation”

An occasional paper series to share ideas, insights and in-depth studies on technology and the
United Nations. The series is sponsored by the Office of Information and Communications
Technology (OICT) but does not necessarily represent the official views of OICT or of the United
Nations.
 Research Objective

Executive Summary
This paper aims to provide an overview of the ethical concerns in artificial intelligence (AI) and the
framework that is needed to mitigate those risks, and to suggest a practical path to ensure the
development and use of AI at the United Nations (UN) aligns with our ethical values. The overview
discusses how AI is an increasingly powerful tool with potential for good, albeit one with a high risk of
negative side-effects that go against fundamental human rights and UN values. It explains the need for
ethical principles for AI aligned with principles for data governance, as data and AI are tightly interwoven.
It explores different ethical frameworks that exist and tools such as assessment lists. It recommends that
the UN develop a framework consisting of ethical principles, architectural standards, assessment
methods, tools and methodologies, and a policy to govern the implementation and adherence to this
framework, accompanied by an education program for staff.

2
 Introduction

Artificial intelligence (AI) has become ubiquitous in our lives—from advertisements that target us as we
browse the web, to autopilot features in cars, airplanes and public transport, to algorithms that screen job
applications; and almost every week, the press informs us with excitement about new applications or
achievements of AI).
The relentless digitization of the world also plays a role. Prompted by remote learning during the
pandemic, a school in Hong Kong now uses an AI to read children's emotions from their video image as
they learn1, in order to assess how well they understood the material. And AIs are able to handle
increasingly sophisticated tasks. Autonomous driving in real-life traffic, for instance, requires a dazzling
range of interpretation and decision making.
Given that AI-based systems have access to and can process infinitely more information than any human,
they are bound to eventually take better decisions than humans. In a number of fields this is already
undeniably true: For example, there are AIs for various forms of cancer detection that outperform the
average trained specialist. That list will continue to grow, and we will increasingly hand over decisions to
AI. The AI being “better” is often defined as faster, more accurate, or more optimal according to certain
criteria. But is that all that matters?
The immediate factors that the AI is programmed to achieve or optimize (example: “get the car from point
A to point B”) cannot not be the only considerations. We can impose constraints (“… and don’t kill any
pedestrians on the way”) but they address only specific problems that may arise. As AIs get increasingly
complex and handle a variety of unpredictable situations autonomously, the only way to ensure the AI
does not do anything “bad”, is to outfit them with a set of general morals and values. We need to build an
ethical framework into our AI.
Stories about advanced AIs creating havoc, even trying to eliminate mankind, have been the topics of
science fiction novels for a long time. While the “AI takes control” scenarios are unlikely for the
foreseeable future, some concerns are quickly becoming reality as AI is increasingly used in financial
systems, law enforcement systems, autonomous cars and weapons.
The UN Secretariat has already started using AI in some forms - Alba, a chatbot available to UN
personnel, is one form of AI. These applications may not seem to be complex enough to need their own
ethical framework, but as we will see further in this paper, some potentially problematic issues are already
manifesting themselves; and in the coming years that need for an ethics framework for AI will become
essential.
And indeed, many governments, regional organizations and businesses have started to consider the way
they are using AI. Many have stated principles, and some have issued policies. The European
Commission, for instance, is one of the main global players in the area of AI policymaking.

1
CNN, 16 Feb 2021. See https://edition.cnn.com/2021/02/16/tech/emotion-recognition-ai-education-spc-intl-
hnk/index.html

3
 From a corporate perspective - other factors may play a role: the desire to be a good corporate citizen
(which, as has been demonstrated, makes a lot of commercial sense too), audibility (what are we doing,
based on which decisions), fiduciary responsibility or compliance.
Apart from preventing undesirable effects of AI, an ethics framework could also outline the positive
contributions an AI should aim to make. From a UN perspective, the potential of AI for Good has been
studied and debated (and implemented) for a number of years2 and it deserves careful consideration to
incorporate in a set of ethical principles.
The aim of this paper is to outline the ethical concerns of AI, now and in the near future, what the UN can
and should do and how to go about it.
This paper is composed of three parts:
Part1: An analysis of the ethical problems with AI: Why should we be concerned? What are the
issues?
Part 2: Defining Ethical AI: A detailed analysis of the ethics of AI, including a stock-take of what has
been done
Part 3: Implementing AI: Proposing a practical way forward for the UN. Includes an overview of
research, examples and current tools
The intention of this paper is that it will lead to discussion followed by action, whether in the form of a set
of guidelines, policy, a code of ethics, or educational initiatives to create awareness of the issues. This
affects not only the technologists who may develop AI-based systems. It also concerns those involved in
procuring or integrating AI systems, those at the managerial level who approve such projects, and
ultimately the users affected by the systems. All these stakeholders need to be aware of the issues and
the measures the UN intends to take, and all need to be taken into account when developing the ethics
framework that guides our development and use of AI.

2
The AI for Good Global Summit organized yearly by ITU is one of the platforms where this is explored. see:
https://aiforgood.itu.int

4
 1. Problems with AI

A. What are the risks of AI?

When an AI system gives a wrong medical diagnosis or the facial recognition in your smartphone
fails, the concern is clear: the system does not accurately perform its task, it makes a mistake. We
can of course wonder to what extent this risk is acceptable. After all, humans also make mistakes.
The captain of the Titanic sailed too close to an iceberg with disastrous results. And at more
personal level, you may have once forgotten your umbrella in the subway or left your dinner on
the stovetop for too long, causing it to burn. We accept some of those mistakes as an inevitable
fact of life, so to what extent should we require an AI to be perfect?

Some of the problems with AI stem from the nature of how the systems are built. An AI is based
on a “model” – a collection of neurons that represent little pieces of knowledge that together form
the algorithm, the functioning of the AI. The knowledge changes over time as the AI “learns” using
feedback from external events, much like a child learns to ride a bicycle. While a computer that
programs itself and improves over time seems great, it also causes concerns. In many cases we
have no idea how successful AI systems function – they are a “black box” that operates in
mysterious ways.

Other problems come from the data that the AI is trained with. It can be useful, sometimes
essential, to provide an AI with large amounts of initial data to learn from, such as, thousands of
X-rays that may contain cancer cells, millions of articles or social media posts on a particular
subject, or hours of video footage of traffic situations. And if that data is tainted or skewed in a
certain way, the AI will mimic that bias. For example, in 2016 Microsoft released Tay, one of the
earliest general purpose chatbots. It was trained using millions of posts on public fora such as
Quora and Reddit – places where anyone can post anything, and which are notorious for
unfiltered content. After some initial success, Tay quickly started using rude and racist language
and had to be taken offline3.

3
https://en.wikipedia.org/wiki/Tay_(bot)
5
 The most sophisticated AI systems need massive amounts of data and computing power. Or to
put it another way, those with access to big data, AI expertise and deep pockets to fund the
computational needs will win in AI. That concentrates the power of this exceptional technology in
the hand of a few companies (which are
What’s a Model?
currently in the United States and in China).
The increase in use of AI in our society risks For an AI to function, it needs to have a
model of the world. That can be done is
exacerbating that imbalance of power, which
several ways, but a popular method is a
will have geopolitical implications. Neural Network. Taking inspiration from the
human brain, a neural network consists of
AI is already being used extensively in the
small entities - artificial neurons - that fire
military. Some nations are already using signals at each other and react to those
autonomous drones that kill humans, and the signals. Together these neurons represent
usage will inevitably increase. An AI arms race the algorithm that determines what the
system does. And this algorithm - the way
might very well be the logical result.
the neurons react- can change over time -
Of course, AI is just a tool, and like any tool, it either because they are manipulated by a
can be used for good or bad purposes. developer or because the AI “learns” from
feedback on its actions. This network of
However, as philosopher Nick Bostrom posited
neurons can have many layers and is
with his Vulnerable World Hypothesis4, AI may typically very large - billions of neurons in
be a technology that allows any group that is some models. Model interpretability -
sufficiently advanced to destroy the world, understanding what layers in the network
represent and why they function the way
regardless of the action of other players. Similar
they do, is then a difficult question.
technologies are nuclear weapons,
biotechnology and, perhaps, nanotechnology.
As such AI requires special attention.

We are currently using what is known as

artificial narrow intelligence (ANI), a form of AI
that is still quite specific to certain tasks. In the
future we may develop artificial general
intelligence (AGI) that reaches a level of
flexibility and general purpose equal to humans
– to be able to react to any situation and solve any task. Some even speculate about artificial
super intelligence (ASI) that far surpasses human cognitive abilities. Most researchers agree that
AGI is decades, if not centuries, away. Of course, even if it is the former, we still need to take an

4
Bostrom, Nick. 2019. “The Vulnerable World Hypothesis.” Global Policy 10 (4): 455–76.
https://doi.org/10.1111/1758-5899.12718.
6
 interest. But it may not help to worry about specifics, since we have so little idea of what will
happen in the meantime and what the world will look like at that point in the future. The best we
can do is build good ethical frameworks for now, and in the future.

B. An overview of the issues

Leaving the geopolitical considerations aside, let us attempt to list the concerns systematically.

1. Incompetence
This means the AI simply failing in its job. The consequences can vary from unintentional
death (a car crash) to an unjust rejection of a loan or job application.

2. Loss of privacy
AI offers the temptation to abuse someone’s personal data, for instance to build a profile of
them to target advertisements more effectively. While this may be an issue related to data, the
nature of AI exacerbates the effects and also makes the situation more complex. For
example, in the United States, a company
called Clearview AI has collected publicly The connection between AI and Data
available photos of users from Facebook, We have seen some of the issues with
LinkedIn and other platforms into a large AI actually pertain to the data it was
database coupled with AI-driven facial trained with. The connection between
recognition algorithms. The company is AI and data is important. Data shapes
selling the system to law enforcement the algorithm any characteristics of the
agencies. In Europe, the General Data data will manifest in the AI. Issues in AI
Protection Regulation (GDPR) would clearly systems are similar to those in data
regulate this form of use of personal data, governance, including data privacy,
but the rules are much less clear when it consent, purpose and intent,
comes to building a profile of a user to target proportionality and protection. When
advertisements, for instance. we establish ethical principles, use
design tools and methodologies or
3. Discrimination monitor and assess operational
systems, we need to look at the data
When AI is not carefully designed, it can and the (AI) algorithm together.
discriminate against certain groups. An
example is the algorithms the company
Palantir developed for the Los Angeles Police Department5. The purpose of the AI was to
support “predictive policing” – predict where crimes will happen so the police can proactively
patrol. The AI was criticized for perpetuating systemic racism as, based on historical data, it
flagged certain neighborhoods as higher risk. More recently, Chinese company Dahua was

5
https://www.technologyreview.com/2020/07/17/1005396/predictive-policing-algorithms-racist-
dismantled-machine-learning-bias-criminal-justice/
7
 criticized for its facial recognition software that can specifically detect people from certain
ethnic backgrounds6.

4. Bias
The AI will only be as good as the data it is trained with. If the data contains bias (and much
data does), then the AI will manifest that bias, too. If the AI is trained with data from a
particular group only, it will start making the wrong assumptions. For example, a recent study
showed that datasets from India drawn from online sources gives a skewed picture, since half
of the population – particularly women and rural residents – do not have access to the
internet.

5. Erosion of Society
AI is used by many platforms that people get their daily news from. In the past, we used to
choose from a range of TV channels and a range of newspapers. Many of us heard or read
the exact same news stories. With online news feeds, both on websites and social media
platforms, the news is now highly personalized for us. We risk losing a shared sense of
reality, a basic solidarity. What makes things worse is that this hyper-personalization is not
just for our benefit (as many platforms claim) – the main goal is to keep us engaged with the
platform. As Tufekci observed in 20187, the AI that powers YouTube found out that people are
drawn to content that is similar to what they started consuming but taken to a further extreme.
In order to keep us glued, the AI suggests evermore extreme content. For example, politically
conservative videos lead to extreme right-wing videos or conspiracy theories, or an interest in
jogging leads to content about ultra-marathons. The AI has no evil intent per se - it is simply
trying to achieve its objective of keeping you engaged.

6. Lack of transparency
The idea of a “black box” making decisions without any explanation, without offering insight in
the process, has a couple of disadvantages: it may fail to gain the trust of its users and it may
fail to meet regulatory standards such as the ability to audit.

7. Deception
AI has become very good at creating fake content. From text to photos, audio and video. The
name “Deep Fake” refers to content that is fake at such a level of complexity that our mind
rules out the possibility that it is fake. We all know that a photo can be altered. But a photo of
a person that is artificially created by an AI is indistinguishable from a real photo. Audio and
video can be faked to sophisticated levels, and we are not used to taking that possibility in to

6
https://en.wikipedia.org/wiki/Dahua_Technology
7
Tufekci, Z. (2018). YouTube, the great Radicalizer. Retrieved March 7, 2021.

https://www.nytimes.com/2018/03/10/opinion/sunday/youtube-politics-radical.html
8
 account. In 2020, hundreds of millions of dollars were stolen by convincing company
employees to make transfers using the deep-faked voice of their CEO.

8. Unintended consequences
Sometimes an AI finds ways to achieve its given goals in ways that are completely different
from what its creators had in mind. Imagine a home robot is tasked to keep the house clean
realizes that the family cat tends to make the
house messy and decides killing the cat is a Should some AI be kept secret?
good way towards its given objective.
Is certain AI technology too
dangerous to make publicly
9. Manipulation available? Should it be restricted,
The 2016 scandal involving Cambridge similar to what is done in the field of
Analytica is the most infamous example where IT security, synthetic biology and
people’s data was crawled from Facebook and nuclear physics? Some researchers
analytics were then provided to target these argue for that.
people with manipulative content for political The opposite argument is that if
purposes. While it may not have been AI per powerful technology is in the hands of
se, it is based on similar data and it is easy to a few only that creates imbalance and
see how AI would make this more effective. increases potential for abuse. OpenAI
was set up with the vision to develop
10. Lethal Autonomous Weapons (LAW) research and make it available to
everyone, thereby leaving the playing
As the use of AI is becoming ubiquitous,
field. However, it ended up
increased use of AI-powered weapons may be
developing advanced algorithms that
inevitable. What is debated as an ethical issue
require such large computational
is the use of LAW — AI-driven weapons that
resources that only a handful of
fully autonomously take actions that
powerful players would benefit from it,
intentionally kill humans. Technologically, the
so they decided not to release it
possibility clearly already exists, whether the
publicly.
final decision (“pulling the trigger”) is left to a
human is a policy choice. The UN has In general, academia plays a large
deliberated the issue8 and a number of groups role in the advances in AI and there is
and some nations have argued for a ban of much openness in that community -
LAW. The UN is the prime forum for such a new techniques are typically shared
discussion9. within months or even weeks.

8
UNGA. “First Committee Weighs Potential Risks of New Technologies as Members Exchange
Views on How to Control Lethal Autonomous Weapons, Cyberattacks“.
https://www.un.org/press/en/2018/gadis3611.doc.htm
9
Gill (2020) “The Role of the United Nations in Addressing Emerging Technologies in the Area of
Lethal Autonomous Weapons Systems” https://www.un.org/en/un-chronicle/role-united-nations-
addressing-emerging-technologies-area-lethal-autonomous-weapons
9
 11. Malicious use of AI
Just as AI can be used in many different fields, it is unfortunately also helpful in perpetrating
digital crimes. AI-supported malware and hacking are already a reality. There is a 2020 report
by EUROPOL, UNICRI and Trend Micro on this topic10.

12. Loss of Autonomy

Respect for personal freedom and autonomy is ingrained into basic human rights in many
societies. Delegating decisions to an AI, especially an AI that is not transparent and not
contestable, may leave people feeling helpless, subjected to the decision power of a machine.

13. Exclusion
The best AI techniques requires a large amount resources: data, computational power and
human AI experts. There is a risk that AI will end up in the hands of a few players, and most
will lose out on its benefits. This would have both societal and geopolitical implications.

Another ethical issue is the Humane treatment of AI. Given that AI is an algorithm, it may seem
odd to worry about its treatment. But as AI becomes increasingly sophisticated and humanoid
robots are developed that look like us and behave like us, the discussion will continue. A number
of recent science fiction movies explore this theme.

10
Trend Micro, Europol, UNICRI. 2020. “Malicious Uses and Abuses of Artificial Intelligence”
http://www.unicri.it/node/3278
10
 2. Defining Ethical AI

C. What is Ethics?
We want our AI to be “good”, or at least to avoid all the negative characteristics we have
identified. We need a comprehensive set of ethical principles that will guide the development and
use of AI. Unfortunately, there is no single truth for ethics.

Ethics are the standards of right and wrong, acceptable and not acceptable, in a certain
community. It includes an individual’s rights, obligations, virtues, the benefits to society, and a
sense of fairness. Ethics will clearly differ for various cultures or groups. What is acceptable in one
country may not be in another. They also change over time. What was commonly accepted 100
years ago is not OK today.

Even for the basic theory of ethics, different approaches exist. Classical philosopher Aristotle
proposed Virtue Ethics (“What kind of person should I be?”). Modern philosopher Immanuel Kant
wrote about Deontological Ethics (“How should I behave?”). Another approach is Consequentialist
Ethics (“What should be my goal?”). Those are only a few of the many other theories of ethics.

Some basic ethical principles go to the core of the questions we are trying to solve about AI. For
instance, Aristotle writes in the Nicomachean Ethics11 about moral responsibility and argues that
“the action must have its origin in the agent”. These ideas become concretely relevant once we
grant AI systems autonomy (that is, agency). Who is responsible for an accident caused by an
autonomous car? The owner? The dealer that sold the car? The manufacturer of the AI? The
supplier of the data that trained the AI?

We can see our ethical values as being

composed of different layers, going from global
to group to individual: Core ethical values
based on inalienable human rights (human
dignity, autonomy); constitutional values (rule of
law, equity, privacy); group specific values
(beliefs or cultural norms); and individual ethical
values (personal convictions).

11
Aristotle. 350 BCE. “Nicomachean Ethics”. Available as free e-book online. See also
https://en.wikipedia.org/wiki/Nicomachean_Ethics”
11
 We can observe the differences in ethics in various countries and cultures in the emphasis they
place in policies and regulations. A recent study shows the differences between individualist and
collectivistic cultures when it comes to moral dilemmas12. Much of the discussion (and regulation)
about data privacy concerns come from countries with highly individualistic cultures. Conversely,
Confucianism teaches that the society is more important than the individual and this is engrained
in the mind of people steeped in this philosophy. At the same time, Buddhism regards the “self” as
an illusion, so one can wonder how relevant the data privacy of the individual then is. That is not
to say there is no concept of a right to privacy or other rights of the individual in these cultures, but
the emphasis is different.

A 2019 study of the global landscape of AI ethics guidelines13 shows that the United States,
European Union and United Kingdom are the most prolific producers of ethics guidelines.
However, there are ethics guidelines being produced in South Korea and Japan as well. In its
2017 AI Plan14, China states the goal to become the world leader in AI by 2030… but also
“emerge as the driving force in defining ethical norms and standards for AI”. Several Chinese
institutions have since developed sets of principles.

Another cultural difference is the balance between protection of human rights and the progress of
business. Let us not forget there is a lot of money at stake in AI. With its General Data Protection
Regulation (GDPR), the EU has implemented some of the most stringent controls on personal
data, but some complain that it stifles business opportunities.

D. What basis do we have?

Agreeing on a set of ethical values is harder if the group of stakeholders is larger and more
diverse. For the UN, the challenge is hardest as we need to serve all people. However, we do
have a basis to work from: The UN Charter, the Universal Declaration of Human Rights (UDHR)
and a number of international treaties give us a basis for AI principles. In fact, some principles
may not need to be reinvented, just applied in the context of AI.

12
Awad et al. 2018. “The Moral Machine Experiment”. https://doi.org/10.1038/s41586-018-0637-6.
13
Jobin, Anna, Marcello Ienca, and Effy Vayena. 2019. “Artificial Intelligence: The Global Landscape
of Ethics Guidelines.” https://arxiv.org/abs/1906.11668.

14
China State Council. 2017. “New Generation Artificial Intelligence Development Plan” (新一代人工智能
发展规划). http://www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm

12
 Ethical values and ideas from all cultures may have useful concepts to contribute to the healthy
principles for AI. The IEEE Standards Association investigated, in its report “Ethically Aligned
Design”15, how applying Buddhist, Ubuntu, and Shinto-inspired ethics could improve responsible
AI. A recent publication from MIT Technology Review explored what Buddhism may offer to the
ethics of AI.16 Some ideas from that paper are that “any ethical use of AI must strive to decrease
pain and suffering” and the “Do no harm principle: the burden of proof would be with those
seeking to show that a particular application of AI does not cause harm”. Another study highlights
ways in which issues surrounding AI in India differ from Western countries and may call for
different approaches to achieve fairness17.

15
IEEE. 2019. “Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous
and Intelligent Systems, First Edition” . https://standards.ieee.org/content/ieee-standards/en/industry-
connections/ec/autonomous-systems.html
16
Soraj Hongladarom / MIT technology Review. 2021. “What Buddhism can do for AI ethics”.
https://www.technologyreview.com/2021/01/06/1015779/what-buddhism-can-do-ai-ethics/

17
Sambasivan et al. 2021. “Re-imagining Algorithmic Fairness in India and Beyond”.
https://research.google/pubs/pub50002/
13
 E. Frameworks for Ethical AI
Many nations, groups or organizations have
published well thought-out sets of AI What’s in a name?
principles that large teams of experts have
A number of terms are used to describe
worked on and each offers its own frameworks of ethical principles for AI. These
perspective. The sets of principles are terms reflect the general mindset, the spirit of the
principles.
entitled with a “brand-name” that gives an
indication of the particular perspective the Ethical AI = ensures compliance with ethical
norms.
group has taken. A bewildering number of
Trustworthy AI = can be trusted by humans as
these go around and they are not always well
acting in a” lawful, ethical and robust” way (AI
defined and sometimes confused: HLEG).
Trustworthy AI; Responsible AI; Explainable Explainable (or explicable) AI, also written as
AI; Interpretable; Transparent; Human- XAI = allows its functioning to be explained to
centered; Safe AI or Mindful AI; and more. stakeholders in non-technical terms. Synonyms
or closely related terms are: Interpretable,
The brand names do not matter nearly as Comprehensible and Understandable AI.
much as the principles that are contained in Interpretable AI = on the same spectrum as
the frameworks; and we see the same Explainability but adds the ability for
stakeholders not only to see but also study the
principles appear in many of them. The box
decision-making process of the AI.
on the right makes an attempt at clarifying
Meaningful AI = used in the Villani report
some of the terms used to describe the (France) to describe a system that is explainable
frameworks, although for most, no single and environmentally friendly, does not increase
definition exists. Below, we will dig deeper exclusion or inequality.

into a few significant ones. Transparent AI = provides some level of

accessibility to the data or algorithm.
The European Commission’s High-Level
Responsible AI = takes into account societal
Expert Group on Artificial Intelligence (AI values, moral and ethical considerations.
HLEG) published one of the best know
Human-centered AI = ensures that human
frameworks in its report “Guidelines for values are central to the way in which AI systems
Trustworthy AI”18. The term Trustworthy are developed, deployed, used and monitored.
here conveys a general sense of trust that Beneficial AI = does not only avoid risks but
humans may have in the AI because of AI contributes positively to society. The notion of
beneficial AI came from one of the Asilomar
that can be trusted by humans to be ”lawful, Principles.
ethical and robust”, where lawful means

18
EC High-Level Expert Group On Artificial Intelligence. 2019. “Guidelines for Trustworthy AI”.
https://ec.europa.eu/digital-single-market/en/high-level-expert-group-artificial-intelligence
14
 comply with the laws in the jurisdiction the AI operates or touches upon, ethical meaning following
a set of ethical principles and robust meaning the AI system is programmed well so as not to
malfunction and cause unintentional harm.

The Montreal Declaration19 is aimed at creating Responsible AI. It contains a sense of societal
responsibility, including the principles of Wellbeing, Solidarity and Democratic participation. It also
mentions Respect for autonomy and Prudence (meaning to be cautious when developing AI,
anticipate any possible negative effects). Accountability and Transparency are other key
elements.

Several others, including the UK Royal Society, use the term Explainable AI. The focus here is
on the ability to explain algorithmic decisions (and sometimes the data driving those decisions is
included) to end users and other stakeholders in non-technical terms. It is the opposite of a “Black
Box”. The AI HLEG describes it as: “auditable and comprehensible and intelligible by human
beings at varying levels of comprehension and expertise”.

There are two types: Model explainability is describing how the AI works on the data as a whole;
however, the model may often be too complex for humans to understand. Sample explainability is
explaining a specific sample — how the AI came to that particular decision.

Explainability is not everything - as the Royal Society states: “Transparency and explainability of
AI methods may […] be only the first step in creating trustworthy systems…”. But explainability is
contained as a principle in many frameworks and there are good arguments for it:

- Validity: Stakeholders need to be reassured of the validity and accuracy of the decisions.
- Auditability: Auditors require tools to check models for regulatory compliance.
- Maintainability: Data scientists need insight in their models to maintain and improve them.
- Transparency: Providing users and other stakeholders insight in the functioning of the AI is useful
to earn their trust.

Some have mused whether Explainability might in the future become mandatory (like GDPR) in
some countries or regions.

Explainable AI has an interesting potential: Often AI (specifically, Machine Learning) is used for
complex environments that humans do not fully understand so we don’t have a model. Weather
patterns or stock price fluctuations are examples. If an AI can learn to predict those complex

19
see: https://www.montrealdeclaration-responsibleai.com/
15
 systems and is built as an Explainable AI, we might learn from the AI to understand the real-world
phenomena better. This is sometime called Third-Wave AI or AI 3.0.

One question that needs to be answered is: “Explainable to whom?” The flip side to explainability
is the protection of intellectual property and strategic advantage. Commercial entities may not
want to go fully explainable if the capabilities of the AI constitute their competitive advantage. And
military users will likely not be keen to explain exactly how the technology in their AI-based
weapons functions.

Organizations that formulate their ethical principles should start by asking themselves what their
main objective is, the general spirit within which they want to work. They are free to use any of the
terms mentioned here or make up their own description. And many do; you will find “Fair AI”,
“Mindful AI”, “Accurate AI”… Remember, what truly matters are the principles.

F. Principles
A large number of frameworks for ethical AI have been developed, each with a number of
principles. We can find similar ideas in individual principles through all frameworks. Even though
the names and the definitions may differ, the basic idea is the same.

The list of principles below is a collected from several frameworks and guidelines including those
of the AI HLEG, IEEE, OECD, Villani report20, ACM, the Japanese Society for AI, the Beijing AI
principles21, the Nation Science and Technology Council (US)22, The Montreal Declaration and the
Asilomar principles, organized in six categories:

20
Villani et al. 2018. “For a meaningful artificial intelligence”. https://

www.aiforhumanity.fr/pdfs/MissionVillani_Report_ENG-VF.pdf.
21
Beijing Academy of Artificial Intelligence. (2019). Beijing AI principles. https://

www.baai.ac.cn/blog/beijing-ai-principles.
22
Holdren / National Science and Technology Council. 2017. “Preparing for the Future of Artificial
Intelligence” https://doi.org/10.1007/s00146-016-0685-0.
16
 A 2020 report in Minds and Machines23 has a more detailed overview of how each of these
principles are represented in each of the frameworks and guidelines. An interesting observation is
that while all principles are represented in some of the guidelines, none of the principles are
represented in all guidelines. However, there seems to be most agreement on the principles of
accountability, privacy and fairness, which are found in 80 percent of the guidelines.

How these principles from the various frameworks address the ethical concerns that we had
identified earlier is shown in the following chart.

23
Hagendorff. 2020. “The Ethics of AI Ethics: An Evaluation of Guidelines”.
https://doi.org/10.1007/s11023-020-09517-8.
17
As we can see, the principles listed under Non-maleficence and Justice contribute to addressing a
range of ethical concerns, while Competence, Accountability and Governance address single, but
important concerns.

18
 G. Governance of Ethical AI policies
So far most ethical principles are used within a limited context — within a particular business or
government agency. General regulation or legislation regarding AI is difficult as traditional
mechanisms can be insufficient for a number of reasons24:

• Machines are not “people” (unless we define them as legal person at some point, just like
businesses are25), and existing legal systems operate by assigning and allocating legal
rights and responsibilities to “persons”.
• The global nature of the internet and the opaque nature of many AIs means that risky AI
development may easily escape (or intentionally evade) detection by regulators.
• AI is often made up of components created by different people at different places (in
different jurisdictions) and times, without conscious coordination (e.g., on public code-
sharing platforms such as Github).
The difficulty in monitoring or enforcing compliance with policies on Ethical AI is a concern that
should be kept in mind when developing principles.

24
Ideas taken from a presentation by Matt Sherer at Asilomar Conference. See:
https://theengineeringofconsciousexperience.com/public-risk-management-for-ai-the-path-forward-matt-
scherer/
25
In 2017, the European Parliament suggested this, but it has not happened yet.

19
 3. Implementing Ethical AI

H. Operationalizing AI principles
Even if it may seem that our use of AI in the UN is still rather limited, and risks can perhaps easily
be contained (probably not the case), it is clear in any case that our use of AI will become
widespread and more complex in the coming years and that the ethical implications are not so
easily foreseen.

Developing a set of ethical principles for AI is the first step. But they will not be realized if there is
no support for those developing, purchasing, integrating, and using AI systems. Stakeholders
need to be involved in the process of defining and operationalizing the Principles – the team that
does this needs to be multi-disciplinary and represent stakeholder communities. It should also
bring in different cultural perspectives (for which the UN, with its multi-national body of staff, is well
positioned). The members of the team will need awareness and understanding of the issues,
development methodologies, standardized assessment methodologies, and technical tools. The
Principles could be provided as a voluntary guideline, a code of conduct or a policy. Sometimes
new policies may not be needed if a reinterpretation of existing policies in the context of AI will
suffice.

Aligning the development of AI systems with Principles can benefit from both technical and non-
technical methods. On the technical side there are architectures, methodology, explanation
methods, and technical tests. Non-technical methods include education, standardization,
stakeholder involvement in design, and diversity in teams.

I. Proposed Framework
For the UN to move toward ethical use of AI in the systems it develops and uses, requires a
number of steps.

Principles: While there are many existing policies and frameworks to take inspiration from, the
UN will need to articulate its own set of principles. These principles should be aligned with the
approach we take to data, including, importantly, the CEB-approved Principles on Personal Data
Protection and Privacy26 and Data Governance policies and guidelines.

Assessment methods: Standard methods should be defined to assess an AI system’s alignment

with our principles. Several methods exist: The AI HLEG report includes a practical assessment

26
See: https://unsceb.org/principles-personal-data-protection-and-privacy-listing
20
 list27, the Z-Inspection process28 is a well thought out method, based on the AI HLEG principles
and the Canadian government has created an “Algorithmic Impact Assessment (AIA)” — an online
questionnaire with 60 questions related to your business process, data and system designed
decisions29.

Assessments can be done both at the start of a project (example: a risk-assessment that will help
decide whether the project should go ahead, and which risks need to be mitigated), before an AI
is deployed or periodically on an AI that is in operation — to ensure they continue to meet the
guidelines.

Architectural Standards: We need minimum standards that specify the technical requirements
AI systems, whether developed in-house or acquired as product or service, must comply with.
Certain principles such as robustness or safety and security would lend themselves particularly to
close governance from an architectural perspective. The architectural standards would take a role
in procurement as well. The WEF “AI Procurement in a box” toolkit30 includes ways to integrate
assessments into a procurement process.

Development Tools and Methodologies: These are meant to enable developers, integrators or
users of AI systems to implement the principles. For instance, methodologies such as Ethics-by-
Design and Value-Sensitive Design (VSD) can help to create AI in a way that leads to more
accountability, responsibility and transparency. A recent paper proposes a modified form of VSD
to use the Sustainable Development Goals (SDGs) as a base for ethical principles for AI that not
only prevent harm but aim to do good31. In terms of tools, several vendors of AI services and
platforms have realized the demand for ethically aligned development of AI and are offering
specific tools. Many open-source tools also exist, such as LIME32, a popular algorithm for AI
explainability.

Awareness and education: This should include different levels. Anyone interacting with AI
systems should be provided with information in non-technical language that explains the principles
and our approach. Developers, data processors and certain users of the AI systems will need
more detailed documentation on our principles including technical specifications as well as

27
See Chapter III of the report: https://ec.europa.eu/digital-single-market/en/high-level-expert-group-
artificial-intelligence
28
Zicari et al. 2021. "Z-Inspection: A Process to Assess Ethical AI".
https://doi.org/10.1109/TTS.2021.3066209. See also: http://z-inspection.org/
29
See:https://www.canada.ca/en/government/system/digital-government/digital-government-
innovations/responsible-use-ai/algorithmic-impact-assessment.html
30
See: https://www.weforum.org/reports/ai-procurement-in-a-box
31
Umbrello and van de Poel. 2020. Mapping value sensitive design onto AI for social good principles.
https://doi.org/10.1007/s43681-021-00038-3
32
Available at https://homes.cs.washington.edu/~marcotcr/blog/lime/. See also an explanation at
https://www.oreilly.com/content/introduction-to-local-interpretable-model-agnostic-explanations-lime/.
21
general education in Ethics and training in specific tools or methodologies. Assessing the
implications of an AI system can be far from trivial and specific education on this will be needed.

Governance: Our Ethical AI principles should result in a policy in the UN entity where it is
implemented. Enforcement of this policy could happen as part of a project approval process, an
architecture review board or it could be linked to data governance. The technically complex nature
of AI systems will often require a balance between oversight (by auditors, senior management or
other bodies) and the engineers developing or installing the AI.

J. Effectiveness
Can we be sure that these efforts by developers and users of AI systems to implement the
principles do indeed lead to AI that is safe, trustworthy, fair or to achieving our desired goals?
Many companies have been criticized give lip-service to AI ethics (“ethics-washing”) while little
actually changes under the hood. But these companies have to balance their adherence to ethics
with commercial interests. The UN, fortunately, does not have this tension.

Still, despite all well-intended efforts, the dividend may not be paid unless we are careful. What is
key is not only to institute a policy, standards, methodologies and tools but importantly to foster a
mindset among staff that both understands and is committed to the spirit of the principles. In that
respect, communication and education must be an important element in our path forward.

Governance Model

22
Conclusion
AI technology has progressed much in the past few years and will continue to do so. Combined
with the fact that our life is increasingly digitized, AI is becoming a powerful tool that will soon be
ubiquitous in all types of devices and services we use in everyday life.
The UN is already using AI and the scope and complexity of its use will likely increase significantly
in the coming years. There are, however, numerous ethical concerns related to the development
and use of AI and we should plan carefully how we want to embrace it. To mitigate negative
effects of an AI, we should establish a framework of principles for ethical AI.
Those principles need to align to the ethical values of the operators, owners, users and other
stakeholders. In the case of the UN, we already have several frameworks to base our principles
on, for instance the UN Charter and the Universal Declaration of Human Rights. Policies and
frameworks have been developed by governments, regional organizations and research groups,
all of which can help inform and inspire our set of principles.
AI is closely connected to data — the information that AI systems are fed with shape the
algorithms. Therefore, ethical principles for AI need to closely align with those used for data,
including the UN Principles on Personal Data Protection and Privacy.
Implementing our ethical principles, that is, ensuring that the AI we develop and use in the future
will meet our ethical standards, requires a policy, standards, methodologies, and tools. But most
importantly it requires awareness and education. The implementation of ethical AI principles all
only succeed if their spirit is understood and adopted by all that interact with AI systems.

23