A High-Level Overview of AI Ethics

Emre Kazim, Adriano Koshiyama

Department of Computer Science, University College London, UK
Contact: [email protected] [email protected]

Abstract
AI ethics is a field that has emerged as a response to the growing concern regarding the impact of
artificial intelligence (AI). It can be read as a subset of the wider field of digital ethics, which addresses
concerns raised by the development and deployment of new digital technologies, such as AI, big data
analytics and blockchain technologies. The principle aim of this article is to provide a high-level
overview of the field as it stands by introducing basic concepts and sketching approaches and central
themes in AI ethics. The first part introduces basic definitions of the terms i.e. defining ‘AI’ and ‘ethics’,
etc.; the second part explores some predecessors to AI ethics, namely engineering ethics, philosophy
of technology and science and technology studies; the third part discusses three current approaches
to AI ethics namely, principles, processes and ethical consciousness; and finally, the fourth part
discusses central themes in translating ethics in to engineering practice. We conclude by summary and
noting future directions and debates.

Key Words: Artificial Intelligence, Governance, Ethics, Philosophy, Regulation

1. Introduction
AI ethics is a field that has emerged as a response to the growing concern regarding the impact of
artificial intelligence (AI). It can be read as a subset of the wider field of digital ethics, which addresses
concerns raised by the development and deployment of new digital technologies, such as AI, big data
analytics and blockchain technologies. There is a growing literature that has significantly increased in
the past number of years (2017 - ) and this literature continues to evolve. The principle aim of this
article is to provide a high-level overview of the field as it stands. We do this by introducing basic
concepts and sketching approaches and central themes in AI ethics.

The overview is structured into four parts; the first part introduces basic definitions of the terms i.e.
what is meant by key terms such as ‘AI’ and ‘ethics’; the second part explores some predecessors to
AI ethics, namely engineering ethics, philosophy of technology and science and technology studies;
the third part discusses three current approaches to AI ethics namely, principles, processes and ethical
consciousness; and finally, the fourth part discusses central themes in translating ethics in to
engineering practice. We conclude by summary of the overview and by noting future directions and
debates.

2. Definitions
In this section we will offer and explicate definitions of key terms. As a point of departure, we begin
by defining first ‘digital ethics’ and then ‘AI ethics’. This ordering follows the fact that AI ethics is a
subdiscipline of the broader umbrella of digital ethics. Following this we then move on to defining how
the term ‘digital’ is being used in digital ethics and then expand on how the term ‘AI’ is being used
(similar to how AI ethics falls under digital ethics, AI is shown to fall under ‘digital’). Further, we
explicate how the term ‘ethics’ is being used in this context and expand upon the dominant ethical
philosophies that AI ethics draws upon. This section closes with an exploration of ‘Human Centric AI’,
which we take to be the overarching value framework of AI ethics (Lukowicz 2019).

2.1 Digital and AI Ethics

Definition: the psychological, social and political impact of emerging digital technologies

Electronic copy available at: https://ssrn.com/abstract=3609292

Where the psychological refers to the likes of agency (moral self-determination), cognitive shifts, and
selfhood; where the social refers to identity, belonging and communities; and, where the political
refers to legal/jurisdictional, democratic (including accountability) and economic. Further, this can be
thought of in terms of the scope of the impact i.e. the psychological represents impact on the
individual, the social the impact on the collective, and the political the impact on the organising
structures of society. Thus, it is also clear that digital ethics is a highly interdisciplinary field, requiring
expertise that spans the sciences (computer science) and the humanities (philosophy, law, sociology
and psychology).

Turing to AI ethics:

Definition: the psychological, social and political impact of AI

This flows from the definition of digital ethics presented above but is specified for AI. In the case of AI
the psychological refers to the likes of mental autonomy, protection from undue manipulation and
the right to know when one is interacting with a non-human agent; the social refers to the likes of
issues of justice and fairness (both procedural and substantive and; the political refers to impacts on
democratic processes and the economy. Like digital ethics, AI ethics is thus highly interdisciplinary
(Kriebitz and Lutge 2020).

Below we further flesh out this working definition by exploring the key constitutive terms, namely
‘digital’ and ‘ethics’.

2.2 Digital
In the context of digital ethics, ‘digital’ is a reference to emerging technologies that are based on
developments in computer science over the past decade.

Broadly construed, the key technologies are:

● Blockchain: decentralised record of digital transactions

● Internet of Things (IoT): any device with an on/off switch that is connected to the internet
● Big Data Analytics: behavioural insight drawn from large information pool
● AI/Machine Learning (ML) and Associated Algorithms: automation of decision making, and
performance of tasks that would normally require human intelligence

Importantly, these technologies are to be thought of together and as interrelated. For example, AI/ML
has been developed and discussed since the mid-twentieth century, however it has grown in
importance and application through advances in computational power and the emergence of large
datasets. As such, big data analytics is a product of AI and big data. Other examples are Federated
Learning, which utilises AI and blockchain (Bonawitz 2019), and the evolution towards Smart Cities ,
which utilse AI, IoT and big data analytics (Burange and Misalkar 2015).

The current epoch is being referred to as the ‘4th industrial revolution’ (Floridi 2014). Where the 1st,
2nd and 3rd industrial revolutions were characterised, respectively, by the use of water and steam
power in the mechanisation of production (circa 1750-1820) , the use of electricity to power mass
production (circa 1870-1920), and the use of electronics and information technologies in the mass
automation of production and processing (circa 1950 -). The 4th is a development upon the third (circa
1990 -), and it is characterized by a fusion of technologies that blur the digital, physical, and biological
spheres (ex. cyberspace, virtual and augmented reality, body-machine interface and robotics)
(Andelka and Mian 2019).

Indeed, two additional themes are i. ubiquitous adoption of these technologies, and ii. futurism.
Where the former is a reference to the increasing use and normalisation of such technologies in
everyday life, government service provision and industry. The latter is a reference to the

Electronic copy available at: https://ssrn.com/abstract=3609292

philosophical/science fictionesque discussions that are emerging as a result of these changes (ex.
debates around the ‘singularity’, transhumanism, and posthumanism (Porter 2017) – often presented
in utopian/dystopia terms). As such, the definition of digital ethics can be expanded and expressed in
terms of the impacts of new digital technologies, through analysis of potential opportunities and risks
in contemporary and future contexts (i.e. it is an applied ethics).

2.3 Artificial Intelligence

As the concern in this overview is with AI ethics, in this subsection we expand upon how we are using
the term ‘AI’ - listed as one of the new digital technologies driving the fourth industrial revolution.

The foundational term here is ‘algorithm’:

Definition: a set of rules or processes that aims to solve a problem or task.

In the digital realm algorithms can be expressed in a computer through programming language. In
order to understand this, we must first note that broadly there are two classes of AI algorithms, which
might be termed: static algorithms – traditional programs that perform a fixed sequence of actions,
usually classified as knowledge-based systems (Schreiber et al. 1993; Giarratano and Riley 1998); and
dynamic algorithms – that learn and evolve by interacting with the environment, usually classified as
ML algorithms (Hastie et al. 2009; Sutton and Barto 2018).

We can think of this as an ‘AI continuum’ of epistemological models (Russell and Norvig 2016), with
the current most successful of which is

ML algorithms - a type of program with the ability to learn without explicit programming, and
can change when exposed to a new environment or information

Traditionally, ML can be broadly subdivided into:

Supervised learning: a program is trained on available and processed data, where specified
inputs are used to predict outputs

Unsupervised learning: the goal of a program is uncover a hidden structure in the data,
thereby ‘discovering’ previously unknown patterns

Reinforcement learning: the goal of the program is to make decisions, achieved through an
iterative trial and error process that is mediated by a reward/penalisation mechanism for
decisions chosen in the development

Machine learning applications in financial services can provide examples of these: Suppose a database
of financial reports is available, if some of them have been historically labelled as positive and
negative, we can leverage this to automatically tag future documents. This can be accomplished by
training an algorithm in a Supervised fashion. If these documents were unstructured, and spotting
relations or topics is the goal (political events, economic data, etc.), an algorithm trained in an
Unsupervised manner can help uncover these hidden structures. Also, these documents can
characterise the current state of the capital markets. Using that, an algorithm can decide which actions
should be taken in order to maximize profits, hedge against certain risks, etc. By interacting and
gaining feedback from the environment (Markets), an algorithm can Reinforce some behaviours so as
to avoid future losses or inaccurate decisions.

In addition to the above mentioned subdivision of ML there are further and disruptive forms of more
advanced ML systems that are making the resolution of previous problems cheaper, faster and more
scalable, like Deep Learning (Goodfellow et al. 2016), Adversarial Learning (Huang et al. 2011),
Transfer and Meta Learning (Andrychowicz et al. 2016; Devlin et al. 2018).

Electronic copy available at: https://ssrn.com/abstract=3609292

2.4 Ethics
Ethics is a broad discipline with considerable scope and plurality of understanding. Although there are
calls and an increasing literature encompassing ethical perspectives of non-Western traditions in AI
ethics literature (for example the IEEE’s Ethically Aligned Design calls for incorporation of non-Western
ethical systems and highlights some of these, including ethics originating in Japan and Africa (2017)),
the predominant discourse is found within the Western European and North American contexts. As
such, whilst bearing in mind the increasing challenge to ethical frameworks representing solely the
‘West’, the ‘ethical’ in AI ethics represents key concepts from the Western philosophical canon. The
definitions below reflect this:

Ethics: the rational and systematic study of the standards of what is right and wrong

Morality: the commonly used term for notions of good and bad

Morality is closely associated with notions of virtue, which concerns the internal state of a person (we
can think of a moral/virtuous act as one that is rooted in the internal state of the said person).

Law: the codified rules and guidelines in a particular jurisdiction

Importantly the law is enforceable i.e. there is a coercive core (usually by the executive branch of the
government). In philosophical ethics we think of this in terms of external constraints, which compares
to the internal states described by notions of virtue.

These concepts are highly related and in natural language use the terms are often interchangeable
and synonymous. Indeed, a law may be considered ethical or unethical; breaking a law may be
considered (im)moral or simply procedural: for example, a person may cross an empty road in a non-
designated place i.e. ‘jaywalk’, however this would not be considered immoral, whereas,
contrastingly, stealing would be a clear example of an immoral breaking of the law. Additionally,
something may be considered immoral, and unethical, but not illegal (ex. infidelity, lying in casual
opinion, etc.). As such, the definitions offered should be understood as notions that allow AI ethics
discussions to be structured rather than stable and inflexible concepts.

2.5 Philosophy of Ethics

The scope of philosophical ethics is vast, with various scholastic schools and traditions, each with its
own community and considerable internal plurality (ex. existentialism, utilitarianism, naturalism,
egoism/hedonism, and deontological/rights ethics, etc.). Two dominant approaches, that can be read
as underpinning common law and continental law, are ‘utilitarianism’ (often referred to as
consequentialism) and rights-based ethics. An addition to these, and less reflected in the context of
contemporary law, is virtue ethics. These are the three dominant ethical theories in academic
philosophy of ethics.

These three are thereby defined below:

Utilitarianism: formulation of principles by considering the consequences of actions that would

result from those rules, where the maximisation of pleasure/minimisation of displeasure is
sought

There are numerous interpretations of utilitarianism (ex. ‘act’ and ‘rule’ utilitarianism) as well as
questions regarding how the terms ‘pleasure’ and ‘displeasure’ are to be understood and quantified.
Putting these concerns aside, the operative concept is that ethics is about weighing the consequences
of actions. One domain where this approach to ethics dominates is that of the justification of
government policy and decision making (ex. policy regarding health care is justified by appeal to
maximization of health outcomes for citizens, economic policy is justified through maximisation of
Gross Domestic Product, etc.).

Electronic copy available at: https://ssrn.com/abstract=3609292

 Rights: entitlements by virtue of belonging to a class

Here, the ethical framework is such that a series of ‘rights’ i.e. that which is referred to as entitlements
in the definition above, are conferred to a person simply by belonging to a class. Where class is
understood as a generic category of identity. Two central examples of rights are human and civil rights.
Human rights are rights entitled to anyone in the class ‘human’, and civil rights are rights
entitled/conferred upon any citizen of a particular jurisdiction. Whereas human rights are considered
inviolable and fundamental simply by virtue of being a human, civil rights are conferred to members
of the political community.

Virtue: development of the character of an individual and actions that result as a consequence
of good character

Virtue ethics (also known as natural ethics) is a classical position that is rooted in pre-enlightenment
Aristotelianism. It is an approach to ethics that emphasises character development – it is closely
associated with ‘perfectionism’, where a person develops over time towards an idealised notion of
the perfect Self (often described in the religious terms of becoming god-like or being in union with the
divine). Good character is understood in terms of values (read, virtues) such as honesty, self-control,
integrity, courage, generosity and fairness.

2.6 Human-Centric AI
In all three ethical approaches, the central subject of concern is the human being i.e. ‘persons’. As
such, it is necessary to offer a working definition of human, which we take to be a rational animal.
There are numerous ways in which this definition can be understood, however, for the present
purpose it is sufficient for us to emphasise that the class ‘human’ is principally defined in terms of
possession of the rational faculty. Thus, humans share all other characteristics with animals
(movement, reproduction, etc.) but are differentiated into a separate class by reason. Reason itself
requires fleshing out and can be thought of in broader terms that include, ‘freedom’, ‘volition’,
‘intentionality’ and ‘agency’. These terms themselves are hotly debated within the philosophical
literature, however for our purposes, we can read them all as referring to reason as an ability to
meaningfully make choices i.e. agency and autonomy.

As such, the human is defined as an ‘agent’ and ‘meaningful choice’ is understood as self-conscious
decision-making. Following from this we can define ‘dignity’ as respect for the moral status of human
beings as rational agents making meaningful choices i.e. existing autonomously. In the context of AI
ethics, autonomy can be subdivided into mental and physical autonomy, where mental autonomy
concerns respect for a person’s deliberative faculties and processes (for example, the right not to be
manipulated consciousness or subconsciously), and, where physical autonomy concerns respect for a
person’s body and choices over their own body.

Furthermore, this aligns with a human-centric approach to AI:

Definition: the development and deployment of AI systems that respect human dignity and
autonomy

Indeed, Human Centric AI can be thought of in positive terms i.e. that automated systems should be
developed and deployed for the betterment of humankind, to advance well-being, human dignity and
human flourishing. When a system reflects this overarching value framework then it can be thought
of as trustworthy i.e. Trustworthy AI – as discussed by the European Commission’s ‘Ethics Guidelines
for Trustworthy AI’ (2019). In section 5.6 below we expand upon themes that fall under
trustworthiness in the context of AI systems.

Electronic copy available at: https://ssrn.com/abstract=3609292

2.7 Conclusion
In the above we have explored philosophical ethics. In the context of AI ethics, the impact upon the
definition offered at the start of the section i.e. the psychological, social and political impact of AI, can
now be understood as a judgement and assessment of these domains through the filter of the ethical
approaches and terminology introduced above. This is a study of normativity, which is the evaluation
and justification of the good. In contrast to this is an anthropological approach to ethics, which is
characterised not by evaluation and rational justification, but instead through study and observation
of what people think and how they behave. It is to learn about ethical behaviour in the world as it is
(Edmond et al. 2018; Arvan 2018). Although we often think of law as the codification of the ethics of
the popular will, most countries do not use direct referenda to legislate. As such there is a complex
relationship between law and morality (as discussed above) and therefore moving directly from
anthropological ethics to codification of law is not standard practice. Instead, anthropological ethics
can be read in terms of gauging respect for democracy which will have consequences concerning trust
in government and democracy, rather than in the straightforward determination of right and wrong
or legality and illegality. One interesting example in the AI ethics literature is to postulate a ‘moral
Turing test’ where a system can be thought of as ethical if it can convince someone interacting with it
that it is reasonably moral, such as how regular human interaction would accommodate ethical
pluralism in human interactions (Allen et al. 2000).

3. Predecessors to AI Ethics
AI ethics is an emergent field that is still in its nascent phase. However, there are a number of
disciplines that have long traditions and literature from which AI ethics draws and can be seen as, in
various ways, a continuation of. The three bodies of literature most relevant are, i. engineering ethics,
ii. philosophy of technology, and iii. science and technology studies. Below we discuss each in turn.

3.1 Engineering Ethics

Definition: the values and ethical system relevant to the practice of engineering

Engineering is a term that refers to structuring, design, and building. It is perhaps the most inherently
practical discipline within the broadly construed sciences (often referred to as an applied science). As
such, it is natural that there already exists a body of literature discussing and debating the social and
environmental impact of engineering. Although engineering encompasses numerous subdisciplines
(such as civil, mechanical, computer and chemical engineering), all of which have specific societal
impacts, the field has matured to the point where non-subdiscipline specific, general community-
based engineering codes have emerged.

For example, the UK based Royal Academy of Engineers, which is a membership by nomination and
selection community of fellows, has a published ‘Engineering Ethics’ guideline ‘Code of Practice’
(2020). The guideline is divided into two parts: first, a ‘Statement of Ethical Principles’ (namely, i.
honesty and integrity, ii. respect for life, law, the environment and public good, iii. accuracy and rigor,
and, iv. leadership and communication); and second, ‘Engineering Ethics in Practice’, which is the
fleshing out of the statement of ethical principles through granular real-world case studies. There are
other similar examples of this (see, IEEE ‘Code of Ethics’ (2020)).

It is noteworthy that Engineering Ethics is predominantly driven by self-assembled communities and

associations, who develop their own standards, are process orientated (ex. case study exploration,
etc), and that they typically go beyond legal compliance.

3.2 Philosophy of Technology

Definition: investigation into the nature of technology and how it impacts the individual, society
and the political

Electronic copy available at: https://ssrn.com/abstract=3609292

There are numerous branches of philosophy that feed into digital ethics. This includes the philosophy
of ethics, as discussed above, but also includes political and social philosophy. The most relevant
philosophical predecessor is the literature referred to collectively as the ‘philosophy of technology’.

The philosophy of technology emerges circa 1920s and can be read as continuing till today (major
figures are Martin Heidegger (d. 1976), Herbert Marcuse (d. 1979), and Jurgen Habermas). It is
differentiated from the philosophy of science, which has a longer legacy in the history of philosophy
and is concerned with method and knowledge. Contrastingly, philosophy of technology emerges as a
result of technological innovation (where technology ‘tekhnē’ means art or craft); indeed, it concerns
the applications and uses of discoveries in science. The principal focus is in appraisal of how technology
affects the human condition and whether the technology is neutral or value laden (ex. exploration of
whether nuclear technology is inherently good or bad, or whether it is dependent on the deployment
of the technology in particular contexts) (Balabanian 2006). From a historical perspective, the
philosophy of technology can be read as a response to the overtly optimistic attitudes of the
enlightenment and ‘positivism’, as well as the Post-WWII world which had an ever-present nuclear
treat and a 1960s-counter culture that was typified by ‘social conscious’. Indeed, it challenges the idea
that there is a necessary connection between scientific discovery and scientific progress, and that this
progress includes and is extended to society. The literature is typically negative, highlighting the
dangers, risks and loss of meaning through adoption of new technologies and increased
technocratisation with respect to the ordering principles of society. Key themes are automation,
alienation, destruction and loss of connection to nature, uniformity, shallow consumption, and
excessive rationalisation.

In addition to noting the general negative critique of technology in the philosophy of technology, it is
also noteworthy that the perspective of the engineers and scientists themselves i.e. the practitioners,
are missing. This negativity also ignores empirical evidence that those who have a less negative view
could cite, namely that technologies such as vaccinations and other medical equipment (ex.
pacemakers etc.) have tangibly and demonstrably decreased mortality rates. This ‘turn to evidence’ is
discussed in the next subsection.

3.3 Science and Technology Studies

Definition: investigation into the effect of culture, society, and politics on scientific
research/activity and technological innovation, and, into the effect of scientific research/activity
and technological innovation on culture, society and politics

The philosophy of technology can be criticised from several points. Two of these are that it is i. typified
by moral panic, and ii. that it is non-empirical. Considering moral panic, namely the phenomena of an
acute reaction to a shocking/dramatic event, typified by knee-jerk reactionaryism - the negative
critique and commentary is read as excessive and blind to the benefits that such technologies have
conferred to humanity. Indeed, it is read as highly politicised, with the philosophy of technology being
instrumentalised as a polemical force of political intervention rather than as a considered investigation
into the nature and impact of technological innovation.

The accusation that the philosophy of technology is non-empirical is motivated by the evolution and
development of the social ‘sciences’, where sciences are placed in quotation marks due to the
increased incorporation and methodological approach of the empirical (natural) sciences in the
humanities (such as sociology and anthropology). Increasingly sophisticated mechanisms of empirical
investigation - that survey, test hypothesis and observe through data analysis (c.f. natural sciences) –
were brought to questions that the philosophers of technology were commentating on.

This empirical turn gave birth to science and technologies studies (circa 1980s -) (Franssen 2016).
Major figures include Bruno Latour and Andrew Light, Donna Haraway. Science and technology studies
problematised the overwhelmingly negative critique of technology by pointing out that technology
7

Electronic copy available at: https://ssrn.com/abstract=3609292

can be designed and used differently and this can lead to radically different social outcomes (ex.
technological developments associated with the internet can be instrumentalised for mass
totalitarian surveillance or instrumentalise to facilitate radical anonymity; similarly, nuclear
technology can be used to make devastating weapons or to produce a steady and reliable source of
energy). In other words, technologies can be used for good and for bad. Evaluation of the value system
that embodies and is present in the deployment of technologies, is done through empirical
investigation. Questions can be investigated empirically regarding who is building the technology, who
is using it, and how has it actually impacted society (and particular groups within society). The notion
that there is a social constructionism/contingency to technology is central to Science and Technology
Studies (Winner 1980; Zittrain 2006). An example of this sociological/empirical turn can be given with
respect to the impact and evaluation of mass consumption (characterised as shallow consumption by
traditional philosophers of technology): surveys/studies conducted showed that people enjoy mass
culture, cinema, and music etc. and that they continue to do so after being presented with the
arguments against the ‘culture industry’ (Curran and Gurevitch 1996). This raised a counter critique
directed towards the philosophers of technology, namely that they fail to respect the aesthetic tastes
and autonomy of non-philosophers and depict people as a gullible mass i.e. patronisation.

3.4 Conclusion
In the previous section we presented a number of definitions, in this section we have presented
predecessor disciples. These definitions and predecessor disciplines will inform what is meant and
what is being drawn upon, in the burgeoning field of AI ethics. In the section below we sketch the
current landscape of AI ethics through a high-level analysis of the main approaches. These are i. the
principles approach, ii. ethical-by-design/processes approach, and iii. ethical consciousness approach.

4. Three Approaches: Principles, Processes and Ethical Consciousness

In the above we defined AI ethics in terms of impact analysis: this can be read as a response to the
increasing number of high-profile cases of harm that has resulted either because of the misuse of the
technology (ex. psychometric voter manipulation (Krotzek 2019), facial recognition surveillance, mass
data collection without consent) or as a result of the technology having design flaws (ex. bias in cases
of recidivism (Chouldechova 2016), loan rejection and medical misdiagnosis) (Bellamy et al. 2018). As
a result, the two main approaches to AI ethics has been a principles approach, which can be read
broadly as an attempt to guide and structure the uses of the technology (thereby mitigating the risk
of misuse) and an ethical-by-design approach, which seeks to mitigate the harms that result from
design flaws. Below we explore these two approaches and also a third, denoted as ‘ethical
consciousness’, which draws from the business ethics literature and concerns a need to institute
particular structures and shifts in cultures, attitudes and norms of those who use, develop and deploy
AI systems.

4.1 Principles
The most vocal response to the harms and risks of new digital technologies, in particular AI, has been
to call for guidelines that inform and direct the use and development of this technologies. Attempts
to delimit these principles fall into three main categories, i. abstract first-principles, ii. development
and application of legislative standards and norms, and iii. to make analogy with bio/medical ethics.
Below we sketch these.

4.1.1 Abstract First-Principles

The first, principles, approach is to articulate a number of statements, typically the expression of a set
of values, and present these as guidance and standards with which the AI systems (and other novel
digital technologies) can be developed and deployed. Such statements of principles have been
produced by all relevant stakeholders, namely academia (Floridi and Cowls 2019), industry (Google:
Artificial Intelligence Principles 2020), NGOs (The Asilomar AI Principles 2017; The Montreal

Electronic copy available at: https://ssrn.com/abstract=3609292

Declaration for Responsible AI 2017), and government (House of Lords Select Committee on
Communications 2018) - although the majority of have been in the context AI, data ethics is heavily
drawn upon.
The principle approach is plagued by several problems (here the examples cited are from the UK House
of Lords report ‘Regulating in a Digital World’ 2018). First, the principles are mostly vague and thereby
difficult to interpret (ex. the accountability principle ‘Individuals and organisations need to be held to
account’, does not provide an expansive explanation of how this would look in practice) (Verma and
Rubin 2018; Dwork et al. 2011). Second, the principles are incongruent i.e. within the same set of
principles, the individual principles contradict one another (ex. both the principle of ‘openness’ and
the principle of ‘privacy’ are asserted together, where respect for one is likely to come at the cost of
implementation and respect for the other). Third, there is considerable lack of clarity regarding
terminology/concepts (ex. principles of accountability, transparency and openness are
complimentary, and in some respects, synonymous expressions). Finally, although there is overlap
between the various statements of principles, there is also a clear lack of consensus. Indeed, to date
(April 2020) there are over 80 AI statements of principles (Jobin et al. 2019; Floridi et al. 2018; Morley
et al. 2019). From an engineering perspective, these problems make it difficult to translate the
principles into practice.

4.1.2 Legislation
The most direct way in which to approach AI ethics is to ensure that the technologies are developed
and deployed in a lawful manner. Indeed, legal compliance is a clear and objective standard by which
to judge and evaluate ethics (where lawfulness can be read as a necessary but insufficient condition
of ethics). However, straightforward recourse to the law is not possible. This is because there are a
number of nuanced concerns. First, the question is posed as to whether it is necessary to create new
laws or to update and ensure application of existing one European Commission 2020). Indeed, one
approach that can be taken is to allow a body of case law to emerge and derive standards, and if need
be legislate accordingly based on this i.e. a ‘bottom-up’ approach (see Committee on Standards in
Public Life 2020). Second, there is the question of whether legislation is appropriate at all, with other
options possible, such as self-regulation and/or a standards body (Simon 2019; Lauterbach 2019).
Third, there is the issue of jurisdiction, where technological development and deployment are
obscured through internationalism, which does not respect or easily lend to jurisdictional oversight
(ex. nation states and international unions (European Union, African Union, etc.)). This also allays itself
the problem of enforcing the law. Fourth, in the common law tradition (typically found in the
anglophone), based on case and precedent, the statutes accommodate ambiguity and contradiction
(relying on judicial discernment). This is a problem with respect to automated systems, where
increasingly there is a call to automate regulatory compliance via expressing laws in codes/protocols
(the ambiguity and contradiction accommodated by common law does not translate in this context).
Regarding this specific concern, continental law (exemplified by EU’s legislative agenda) typified
through a top-down (first principles) legal philosophy is less likely to have this problem. Finally, and
more specifically, in the context of AI there is the question regarding the legal status of an algorithm
(ex. will algorithms follow how companies have rights and obligations, and will AI systems will have
artificial personhood status) (Treleaven et al. 2019), and how the questions of legal culpability, which
rely on judgements of intent, can be formulated in the context of AI systems (Rajakishore and Sahu
2020; Vetrò et al. 2019).

4.1.3 Bio/Medical Ethics

Bio/medical ethics can be turned to for inspiration when approaching and developing AI ethics due to
the fact that it is well established, robust, has accountability mechanisms and is an example of the
ethics of an applied science with significant social impact (Mittelstadt and Floridi 2016; Mittelstadt
2017; Christine 2019); this is not to be confused with applications of new technologies in medicine
(Char et al. 2018; Nuffield Council on BioEthics 2018) or the ethics of using AI in medicine (Lamanna

Electronic copy available at: https://ssrn.com/abstract=3609292

and Byrne 2018). However there are important disanalogies between bio/medical ethics and any AI
ethics scheme that may emerge. In the context of AI systems there is: first, no common aims and
fiduciary duties (legal or ethical relationships of trust) i.e. the relationship between the doctor and the
patient is disanalogous to the relationship between the engineer/company and the public); second,
there is no professional history and norms i.e. the technologies are still at an early stage and the field
of AI ethics is still contested (as discussed above); and finally, no robust legal and professional
accountability mechanisms is present i.e. in medicine a doctor can be ‘struck off’ and/or a licence
withdrawn (Mittelstadt 2019).

4.2 Processes
A second approach to AI ethics is to address risk and harm that can result because of design issues and
lack of appropriate governance.
4.2.1 Ethical-by-Design
An ethical by design approach is a commitment to building systems ethically and, in the hope, that
harm can be prevented. There are several approaches to ethical-by-design. First, via co-design, which
is a reference to interdisciplinarity in the design processes. The idea here is that AI engineers may not
be best placed to understand and discern the ethical dimension and potential impact of the
technology and as such experts from anthropology, sociology, philosophy, psychology, law, etc. i.e.
‘ethicists’, can be integrated into the team at the development stage. Second, by having clear
principles, laws, standards and guidelines, with which to structure and judge design. As noted above,
there is considerable ambiguity and lack of consensus in the field of digital regulation (perhaps with
the exception of GDPR) and standards, thus making it difficult to establish best practices in the domain
of translating principles into engineering practice. Finally, and as a corollary to the previous point,
from a design perspective, implementation of ethical principles will need to be balanced and ‘traded-
off’. For example, in the design, emphasis on transparency and openness may come at the cost of
privacy. These judgements need to be justified, articulated and expressed in different forms in various
contexts. With this practical dimension and necessity of trade-offs in mind, several practical manuals
have been published, all of which note these trade-offs and discuss ethical evaluation (including
interdisciplinarity) in the design, development and deployment phases, respectively (IEEE 2017;
Whittlestone et al.; 2019; Leslie 2019).

4.2.2 Governance
Within the processes approach to AI ethics questions of governance are emerging. More generally,
with respect to novel digital technologies governance can be divided into two broad streams, namely
technical and non-technical (with ethical-by-design falling into the former category). Technical
governance concerns systems and processes that render the activity of the technology itself
accountable and transparent – this includes justifying what design choices are made and ensuring the
system is accessible. Non-technical governance concerns systems and processes that focus on
allocating decision makers, providing appropriate training and education (in the context of new digital
technologies such as AI, education and training will require continuous updating), and keeping the
human-in-the-loop with respect to how automated decisions are used while respecting human rights
(often referred to in the context of ‘human centric AI’ (Lukowicz 2019).

Falling under governance is the growing literature on auditing and impact assessments. Auditing and
impact assessments involve the creation of metrics for tracing and tracking decisions, making the
technologies accessible for verification and accountability. The most well-established form of this is
Data Protection Impact Assessments (DPIA) (Information Commissioner’s Office 2019). To date more
general ‘data ethics’ canvases and process frameworks have emerged (DCMS 2018; Open Data
Institute 2020), and impact assessments specific to AI are being called for and developed (European
Commission 2020; Jordan et al. 2019; Babuta and Oswald 2020).

10

Electronic copy available at: https://ssrn.com/abstract=3609292

4.3 Ethical Consciousness
Ethical consciousness refers to a person or institution or cultural norm, that has a disposition that is
motivated by a moral awareness rather than, say, exclusively a concern of economics (pay and profit),
or legality (responsibility, culpability and compliance). In other words, this is a desire to ‘do-the-right
thing’. Ethical consciousness can be read as coming out of business ethics (Lutge 2020; Weiner 2019),
which is an applied ethics within the commercial environment. Sharing many of the themes from the
previous section, it encompasses the integration of codes of conduct and compliance, however it also
expands to consider reputational issues, (corporate) social responsibility, and, most relevant to the
development of ethical consciousness, concerns for institutional philosophy and culture. Drawing
particularly on the latter, ethical consciousness can be stated in terms of societal and culture shifts in
the awareness of citizens, technology developers and deployers, policy makers and leaders of
industry, in the ethical dimensions of new digital technologies. Such a shift will be facilitated through
an increase in digital literacy, particularly important for meaningful human intervention and issues of
consent (Miller and Coldicutt 2019).

4.4 Conclusion
The above approaches can be thought of in terms of i. the theoretical and abstract, ii. the practical
and process, and iii. culture and society, and in turn can be thought of as all necessary in the
development of a mature AI ethics i.e. one that is truly reflective of the nuances and the inherent
complexity found in all forms of applied ethics. This also challenges any attempt to silo questions and
responsibilities; AI ethics is not exclusively in the purview of philosophers or lawyers or sociologists or
engineers, etc. rather it is inherently interdisciplinary. As such, the call for interdisciplinary must be
followed with development of methods and structures by which they can be fulfilled. It is likely that
this aligns with the call for training and education (noted above in the context of discussions regarding
governance); as such an increase in ‘literacy’ is crucial in informing all relevant stakeholders and
society at large and will be facilitated through a holistic education and training agenda (touching upon
and integrating ethics, policy and engineering).

5. Major Themes in AI Ethics

There are many terms and phrases that have emerged within the AI ethics literature. For example, a
comprehensive (2019) review of AI ethics guidelines found eleven ‘ethical principles’ namely, 1.
Transparency, 2. Justice, fairness, equity; 3. Non-Maleficence 4. Responsibility and accountability; 5.
Privacy; 6. Beneficence; 7. Freedom and autonomy; 8. Trust; 9. Dignity; 10. Sustainability; and 11.
Solidarity. These principles were identified through frequency of the terms (and their synonyms) in
the literature (the terms above are listed in order of prevalence). However, as noted above, there is
overlap and these terms require considerable disambiguation (Jobin et al. 2019). For the purposes of
this overview we draw on the growing engineering expertise that overlaps with the ethics principle
space. Indeed, in the section below, we identify and explore six themes that we believe encompasses
the attempt to bridge the need to implement ethics into engineering and systems. These are, human
agency and oversight, safety, privacy, transparency, fairness, and accountability. Drawing on the
European Commission’s ‘Ethics Guidelines for Trustworthy AI’ (2019), we note that these themes can
be read as falling under the umbrella of ‘Trustworthy AI’ (introduced above in section 2.6).

5.2 Human Wellbeing

This theme is grounded on the ethical principle of respect for human dignity (see section 2.6) and
includes psychological, social, and environmental wellbeing. Here key themes are:

● Impact on human agency: This touches on the impact on individuals and, in particular, mental
autonomy. For example, consider whether a system directly or indirectly diminishes the
deliberative/rational capacity of humans (ex. cognitive shifts in attention spans). Another
issue is consent, where respect for human agency would entail meaningful and informed

11

Electronic copy available at: https://ssrn.com/abstract=3609292

 consent including the right to withdraw consent and be presented with consent mechanisms
that are explicable in the context of an average user.
● Societal Impact: The societal refers to identity, belonging and communities, and includes the
political legal/jurisdictional, democratic and economic impacts. Citizen rights fall here, where
issues of fairness – both procedural and substantive, and bias should address. With respect to
the economic, concerns include, fair competition as well as setting the framework for
competition (Khan et al. 2016). Finally, the environmental impact should be considered,
including issues of sustainability.

The above may be read as a re-statement of the ethical imperative that grounds Human Centric AI,
i.e. that automated systems should be developed and deployed for the betterment of humankind, to
advance well-being (or at least not adversely affect it), human dignity and human flourishing. Indeed,
although the discussion is presented in terms of mitigating risks and potential harms, it is important
to bear in mind the considerable benefits of AI for people and society (Taddeo and Floridi 2018).

5.3 Safety
This theme is based on the ethical principle of preventing harm, where harm is defined in terms of
adverse effects on human well being i.e. the psychological, social and environmental human
wellbeing.
Here the approach is one of identifying risks and then mitigating for them: crucially the approach is
preventative. Key themes here are:

● Robustness: Systems should be robust against adversarial attacks i.e. hacking. Here resilience
is important and that there are measures to stop/resist exploitation of a system (ex. data
poisoning, model leakage).
● Malicious use: A system may have been developed for one use and then be appropriated
and/or modified for another, malicious, use i.e. dual use (ex. the weaponization of delivery
drones).
● Reliability and Reproducibility: Reliability concerns the system working within the framework
of why it was developed and deployed, whereas reproducibility concerns consistent behaviour
when given the same set of inputs and under the same conditions. In the context of robustness
this is important because a system that is unreliable and does not reproduce results will lead
to untrustworthiness in the system.
● Fallback Plans and Unknown risks: A concern for robustness is to address known risks (such as
those cited above, i.e. security, malicious use, reliability and reproducibility) and unknown
risks. With respect to the former, safeguards can be put in that specifically monitor and track
usage and/or metrics of known risks and put in place stops or other mechanisms that would
mitigate this risk. With respect to the latter, it is not possible to fully anticipate risk and as
such, mechanisms can be put into place to mitigate this (ex. fallback mechanisms, automatic
stops – statistical or rule-based, metrics, periodic request for human operators to continue
operating, etc).

5.4 Privacy
This theme is based on the public and political demand to respect a human’s personal information.
This relies on a distinction between the private-personal and the public-political/communal sphere,
where the former is seen as demanding a higher level of respect for privacy than the latter (Mittelstadt
2017). Informed consent is crucial here, where people are informed and updated regarding the
storage and use of their data. Further, there are debates concerning the value derived from personal
data and the distribution of financial benefits derived thereof (ex. mass data driven business models).
Additionally, privacy has emerged as a political concern, with mass surveillance and personal data
being used to target and engage in recommendation and manipulation (for both political and
economic ends).

12

Electronic copy available at: https://ssrn.com/abstract=3609292

 ● Data Stewardship: The management of data spans many stages, including collection, pre-
processing, tracking providence, analysis, publication of results, re-use and recycling, all while
maintaining security and where appropriate anonymising. Stewardship is the management of
this multi-layered process and has come to be a discipline and set of skills in its own. Part of
this remit is data protection, which is crucial to preserving privacy with respect to who has
access to the data (in particular personal data).
● Data Minimisation: Within the context of privacy and data protection, a generalised principle
to use only the amount of data that is needed is referred to as data minimisation. Here, three
dimensions are identified, namely, i. adequacy: where the data is sufficient to fulfil a stated
purpose; ii. relevant: where the data has a justifiable link the stated purpose, and; iii.
necessary: where the data is limited and no more than what is needed is held (and where
appropriate, deleted when no-longer in use for the stated purpose).

5.5 Transparency
This theme is based on the principle of openness, which is crucial to establishing trust and
accountability. Transparency can be thought of with respect to what decisions are being made
regarding how the AI system is used and with respect to how the system comes to its decisions. The
former touches on governance (which is also expanded upon in the accountability section below),
whereas the latter concerns explainability of automated decision makers. As such key themes are:

● Explainability: Being able to explain how a system has come to a decision (c.f. black-box
problem) and making that decision explicable to various stakeholders i.e. explicability will
depend on the technical knowledge of the person, what role they play in the development
and deployment of the system and what kind of end-user they are. Furthermore, there are a
host of technical requirements and tools that may be grouped under explainability, namely,
accuracy, traceability, tracking, general (global/model) and specific (local/data point)
explanations.
● Communication: In addition to explicability i.e. communication, of automated decisions, there
is also the concern for communicating the capabilities and purposes of the system to those
both directly and indirectly impacted (Binns 2018). One crucial dimension is that in cases
where a system may be mimicking human subjectivity (ex. a chatbot) it should be
communicated clearly that the user is interacting with an AI system.

5.6 Fairness
This theme is based on the ethical principle of human equality. Fairness falls under debates about
justice and is hotly contended. A central question is what definition(s) of fairness/justice to commit to
i.e. there are mutually exclusive theories of fairness such as corrective, distributive, procedural,
substantive, comparative, … etc. The question is also raised as to the scope or remit within which
notions of fairness/justice are being discussed i.e. fairness in the context of political communities
(citizenship rights) and/or universal human concerns, and, if appropriate, how to define demographics
i.e. gender, nationality, race, socio-economic background, … etc. Key themes here are:

● Bias: Here bias refers to preferential or discriminatory treatment of persons or groups.

Concerns touch upon bias in (historical) datasets, intentional exploitation of people (ex.
customers/regional pricing) and quality of service provision. This also includes a distinction
between fairness in terms of treatment and fairness in terms of impact (Lipton et al. 2018).
● Accessibility: Although much of the discourse in AI ethics concerns mitigation of harms, it is
also clear that there are significant benefits that people and society will gain from these
systems. As such it is paramount that all people, to the greatest extent possible, have equal
access to these technologies; aside from affordability, designs should be user-friendly (ex.
towards different demographics, cultural and linguistic groups, and in particular, those with
disabilities) i.e. there is not a not-size-fits-all approach.
● Participation: Communication, in an accessible and explicable vernacular, to users will
facilitate meaningful engagement of wider society with AI systems. This will also facilitate

13

Electronic copy available at: https://ssrn.com/abstract=3609292

 learning and develop a more holistic approach to consent. Participation also includes,
soliciting the views of stakeholders during the development of the system. This expands to
diversity (in option and background) in hiring and the interdisciplinary teams involved in
governance and development.

5.7 Accountability
Ethical AI is a branch of applied ethics and, as such, is inherently concerned with how AI systems
impact human beings. How the systems are developed, the processes, logic of decision making, the
allocation of duties with respect to who makes decisions, and how and to what extent, where impacts,
risks and harms gauged and measured. All of this falls under the remit of accountability, and, as the
previous sentence indicates accountability relates to knowing who had made decisions, how those
decisions were made and what systems or tools were put in place to measure and track i.e.
governance. Finally, accountability is central for the possibility of redress and assigning legal liability.
Key themes here are:

● Keeping it Human
Crucial to accountability is ensuring that there are robust human oversight mechanisms, this is based
on the principle, and current legal standing (see section 4.1), that humans are ultimately the
accountable and thereby responsible for harms that may result from AI systems. Within the literature,
there is a growing discourse regarding keeping the ‘human-in-the-loop’ (Wang 2019), which is
discussed in two ways, firstly as ensuring that there is human intervention in the decision processes
of automated systems and secondly in terms of human oversight regarding automated decisions i.e.
in the context of ‘decision support systems’. With respect to the latter, a ‘semi-automated decision’
scheme can be thought of, where a system generates results, directions and recommendations (ex.
whether to hire someone, or reject a loan application) and is followed by human review in order to
affirm or reject the recommendation (Information Commissioner’s Office 2020). In these cases, checks
should be in place to assure that the human review does not become a rubber stamp exercise
rendering the decisions effectively solely automated. Ensuring human responsibility also entails
mapping of duties and risks to responsibilities and roles within an institution.
● Algorithmic Impact Assessments: This relates to direct mechanisms by which to assess and
thereby put in measures to mitigate potential harms of AI systems. We can divide this into
two, namely, i. impact assessments and ii. auditing of technology (Ada-Lovelace Institute and
DataKind UK 2020).
i. Impact Assessments can range from assessments of fundamental rights,
psychological and social well being (ex. social cohesion), citizen rights,
democracy, economic and environmental impacts.
ii. Auditing can be directed to the technology itself, and focus on fairness (ex.
tracking bias metrics), explainability (ex. providing global and local
explanations of models and individual decisions) and robustness (ex. testing
how resilient a system is to hacking).
It is important that these are conducted in such a way as to facilitate inspection (perhaps even
independently conducted). Moreover, clear documentation is necessary. This includes
documenting any trade-offs and the methodology and logic behind trade-off choices.

6. Conclusion
In this high-level overview we have introduced basic definitions of terms such as ‘AI’ and ‘ethics’.
Following this we explored some predecessors to AI ethics, namely engineering ethics, philosophy of
technology and science and technology studies/ We then discussed three current approaches to AI
ethics namely, principles, processes and ethical consciousness. Turning to translating AI ethics into
engineering practice we surveyed the themes of human centric AI, safety, transparency, fairness and
privacy.
Themes and developments in AI ethics that we anticipate:

14

Electronic copy available at: https://ssrn.com/abstract=3609292

 ● Data Ethics and AI: given the substantial literature, practice and regulation around data ethics,
we anticipate that the relationship between data ethics and AI will increase in importance.
This includes whether the two are compatible, whether one is prioritised over the over i.e.
will AI ethics ‘sit’ on top of data ethics or will data ethics have to be reconsidered and
reformulated in light of increased AI adoption? We anticipate that this debate will be both
conceptually important and have significant regulatory/practical consequences.
● Legal Status of Algorithms: raised in section 4.1.2, the legal status of an algorithm, with
respect to responsibilities and obligations of those developing and deploying them, is likely to
raise a number of complex questions regarding the nature of legal culpability and even
questions of agency and personhood. We anticipate that this question will increase in
complexity and importance the more AI systems are embedded in people's daily lives and in
proportion to the function of these systems (ex. sectors such as medicine may require nuances
that other sectors such as entertainment will not).
● AI and the Economy: we believe that the relationship between AI and the economy will
become a major theme of AI ethics. In addition to the current discussions of automation and
the loss of labour, which allay into questions such as universal basic income, etc., there are
broader questions regarding taxation of AI systems, national and international procurement
standards and strategies, and the strategic importance of AI in national budgets. Naturally the
economic conditions will give rise to political implications and as such, beyond concerns for
misuse of AI systems in the democratic process (ex. voter manipulation), debates about how
AI impacts the structure of the state, the very notion of a nation (with clear juridical remit)
and trust in government communication, management and service provision, will become
central themes within AI ethics.

7. References
1. Allen C, Varner G, & Zinser J. (2000). Prolegomena to any future artificial moral agent.
Journal of Experimental & Theoretical Artificial Intelligence, 12(3), 251-261.
2. Andrychowicz M, Denil M, Gomez S., Hoffman MW, Pfau D, Schaul T, & De Freitas N. (2016).
Learning to learn by gradient descent by gradient descent. In Advances in neural information
processing systems (pp. 3981-3989).
3. Artificial Intelligence (AI) in healthcare and Research. (2018). Nuffield Council on BioEthics.
4. Artificial Intelligence and Public Standards: report. (2020). Committee on Standards in Public
Life, Chair, Lord Evans of Weardale KCB DL.
5. Arvan M. (2018). Mental time-travel, semantic flexibility, and AI ethics. AI & SOCIETY, 1-20.
6. Awad E, Dsouza S, Kim R, Schulz K, Henrich J, Shariff A, & Rahwan I. (2018). The moral
machine experiment. Nature, 563(7729), 59-64.
7. Babuta A, Oswald M. (2020). Data Analytics and Algorithms in Policing in England and Wales:
Towards A New Policy Framework. Royal United Services Institute.
8. Balabanian N. (2006). On the presumed neutrality of technology. IEEE Technology and
Society Magazine, 25(4), 15-25.
9. Bellamy RK, Dey K, Hind M, Hoffman SC., Houde S, Kannan K, ... & Nagar S. (2018). AI
Fairness 360: An extensible toolkit for detecting, understanding, and mitigating unwanted
algorithmic bias. arXiv preprint arXiv:1810.01943.
10. Binns R. (2018). Algorithmic accountability and public reason. Philosophy & Technology,
31(4), 543-556.
11. Bonawitz K, Eichner H, Grieskamp W, Huba D, Ingerman A, Ivanov V, & Van Overveldt T.
(2019). Towards federated learning at scale: System design. arXiv preprint arXiv:1902.01046.
12. Burange AW, & Misalkar HD. (2015, March). Review of Internet of Things in development of
smart cities with data management & privacy. In 2015 International Conference on Advances
in Computer Engineering and Applications (pp. 189-195). IEEE.
13. Char DS, Shah NH, & Magnus D. (2018). Implementing machine learning in health care—
addressing ethical challenges. The New England journal of medicine, 378(11), 981.

15

Electronic copy available at: https://ssrn.com/abstract=3609292

14. Chouldechova A. (2016). Fair prediction with disparate impact: A study of bias in recidivism
prediction instruments, arXiv pre-print.
15. Christine H. (2019). Making the Ethical Practical: Ideas for Applying a Digital Health Ethical
Framework for Charities. DataKind.org https://www.datakind.org/blog/making-the-ethical-
practical-ideas-for-applying-a-digital-health-ethical-framework-for-charities.
16. Curran J. and Gurevitch M. (1996). Mass media and society. 2nd Edition, London, UK:
Edward Arnold, 305 – 324.
17. Data Ethics Canvas. (2020). Open Data Institute https://theodi.org/article/data-ethics-
canvas/
18. Data Ethics Framework. (2018). Department for Digital, Culture, Media and Sport
19. Devlin J, Chang MW, Lee K, & Toutanova K. (2018). Bert: Pre-training of deep bidirectional
transformers for language understanding. arXiv preprint arXiv:1810.04805.
20. Dwork C, Hardt M, Pitassi T, Reingold O, & Zemel RS. (2011). Fairness Through Awareness.
CoRR abs/1104.3913 (2011). arXiv preprint arXiv:1104.3913.
21. Engineering ethics (2020) Royal Academy of Engineering
(https://www.raeng.org.uk/policy/engineering-ethics/ethics (accessed 24 April 2020)
22. Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and
Intelligent Systems, Version 2. (2017). IEEE.
23. Ethics Guidelines for Trustworthy AI: High-level Expert Group On Artificial Intelligence’ (8th
April 2019) European Commission
24. Examining the Black Box: Tools for assessing algorithmic systems. (2020). Ada-Lovelace
Institute & DataKind UK. https://www.adalovelaceinstitute.org/examining-the-black-box-
tools-for-assessing-algorithmic-systems/.
25. Floridi L, & Cowls J. (2019). A unified framework of five principles for AI in society. Harvard
Data Science Review.
26. Floridi L, Cowls J, Beltrametti M, Chatila R, Chazerand P, Dignum V, & Schafer B. (2018).
AI4People—An ethical framework for a good AI society: Opportunities, risks, principles, and
recommendations. Minds and Machines, 28(4), 689-707.
27. Floridi L. (2014) The 4th Revolution. Oxford University Press
28. Franssen M, Vermaas PE, Kroes P, Meijers AWM. (Eds.) Philosophy of Technology after the
Empirical Turn. Springer (2016).
29. Giarratano JC, & Riley G. (1998). Expert systems. PWS publishing co.
30. Goodfellow I, Bengio Y, & Courville A. (2016). Deep learning. MIT press.
31. Google: Artificial Intelligence Principles. (2020). https://ai.google/principles/
32. Guidance on the AI auditing framework: Draft guidance for consultation. (2020) Information
Commissioner’s Office.
33. Guide to the General Data Protection Regulation (GDPR). (2019). Information
Commissioner’s Office. https://ico.org.uk/media/for-organisations/guide-to-data-
protection/guide-to-the-general-data-protection-regulation-gdpr-1-0.pdf
34. Huang L, Joseph AD, Nelson B, Rubinstein BI, & Tygar JD. (2011, October). Adversarial
machine learning. In Proceedings of the 4th ACM workshop on Security and artificial
intelligence (pp. 43-58). ACM.
35. IEEE Code of Ethics. (2020). IEEE, https://www.ieee.org/about/corporate/governance/p7-
8.html)
36. Jobin A, Ienca M, & Vayena E. (2019). The global landscape of AI ethics guidelines. Nature
Machine Intelligence, 1(9), 389-399.
37. Jordan S, Fazelpour S, Koshiyama A, Kueper J, DeChant C, Leong B, & Shank C. (2019).
Creating a Tool to Reproducibly Estimate the Ethical Impact of Artificial Intelligence.
38. Khan AQ, Khwaja AI, & Olken BA. (2016). Tax farming redux: Experimental evidence on
performance pay for tax collectors. The Quarterly Journal of Economics, 131(1), 219-271.
39. Kriebitz A, & Lütge C. (2020). Artificial Intelligence and Human Rights: A Business Ethical
Assessment. Business and Human Rights Journal, 5(1), 84-104.

16

Electronic copy available at: https://ssrn.com/abstract=3609292

40. Krotzek JL. (2019). Inside the Voter’s Mind: The Effect of Psychometric Microtargeting on
Feelings Toward and Propensity to Vote for a Candidate. International Journal of
Communication, 13, 3609–3629.
41. Lamanna C, & Byrne L. (2018). Should artificial intelligence augment medical decision
making? The case for an autonomy algorithm. AMA journal of ethics, 20(9), 902-910.
42. Lauterbach A. (2019). Artificial intelligence and policy: quo vadis?. Digital Policy, Regulation
and Governance.
43. Leslie, D. (2019). Understanding artificial intelligence ethics and safety: A guide for the
responsible design and implementation of AI systems in the public sector. The Alan Turing
Institute.
44. Lipton Z, McAuley J, & Chouldechova A. (2018). Does mitigating ML's impact disparity
require treatment disparity?. In Advances in Neural Information Processing Systems (pp.
8125-8135).
45. Lukowicz P. (2019). The Challenge of Human Centric AI. Digitale Welt 3, 9–10 (2019).
46. Miller C, & Coldicutt R. (2019). People, Power and Technology: The Tech Workers’ View.
Doteveryone https://doteveryone.org.uk/report/workersview.
47. Mittelstadt B, & Floridi L. (2016). The ethics of big data: current and foreseeable issues in
biomedical contexts. Science and engineering ethics, 22(2), 303-341.
48. Mittelstadt B. (2017). Designing the health-related internet of things: ethical principles and
guidelines. Information, 8(3), 77.
49. Mittelstadt B. (2017). From individual to group privacy in big data analytics. Philosophy &
Technology, 30(4), 475-494.
50. Mittelstadt B. (2019). Principles alone cannot guarantee ethical AI. Nature Machine
Intelligence, 1-7.
51. Morley J, Floridi L, Kinsey L, & Elhalal A. (2019). From What to How: An Initial Review of
Publicly Available AI Ethics Tools, Methods and Research to Translate Principles into
Practices. Science and Engineering Ethics, 1-28.
52. Nath R, & Sahu V. (2020). The problem of machine ethics in artificial intelligence. AI &
SOCIETY, 35(1), 103-111.
53. Phillips AM, & Mian IS. (2019). Governance and Assessment of Future Spaces: A Discussion
of Some Issues Raised by the Possibilities of Human–Machine Mergers. Development, 62(1-
4), 66-80.
54. Porter A. (2017, June). Bioethics and transhumanism. The Journal of Medicine and
Philosophy: A Forum for Bioethics and Philosophy of Medicine, 42 (3), 237-260. Oxford
University Press.
55. Regulating in a Digital World. (2019). House of Lords select committee on communications
https://publications.parliament.uk/pa/ld201719/ldselect/ldcomuni/299/299.pdf
56. Russell SJ, & Norvig P. (2016). Artificial intelligence: a modern approach. Malaysia; Pearson
Education Limited.Hastie T., Tibshirani R., & Friedman J. (2009). The elements of statistical
learning: data mining, inference, and prediction. Springer Science & Business Media.
57. Schreiber G, Wielinga B, & Breuker J. (Eds.). (1993). KADS: A principled approach to
knowledge-based system development (Vol. 11). Academic Press.
58. Simon, JP (2019). Artificial intelligence: scope, players, markets and geography. Digital Policy,
Regulation and Governance.
59. Sutton RS, & Barto AG. (2018). Reinforcement learning: An introduction. MIT press.
60. Taddeo M, & Floridi L. (2018). How AI can be a force for good. Science, 361(6404), 751-752.
61. The Asilomar AI Principles. (2017). https://futureoflife.org/ai-principles/?cn-reloaded=1
62. The Montreal Declaration for Responsible AI. (2017). https://www.montrealdeclaration-
responsibleai.com/ (accessed 24 April 2020)
63. Treleaven P, Barnett J, & Koshiyama A. (2019). Algorithms: law and regulation. Computer,
52(2), 32-40.
64. Understanding artificial intelligence ethics and safely. (2019). Government Digital Service
and Office for Artificial Intelligence.

17

Electronic copy available at: https://ssrn.com/abstract=3609292

65. Verma S, & Rubin J. (2018, May). Fairness definitions explained. In 2018 IEEE/ACM
International Workshop on Software Fairness (FairWare) (pp. 1-7). IEEE.
66. Vetrò A, Santangelo A, Beretta E, & De Martin JC. (2019). AI: from rational agents to socially
responsible agents. Digital Policy, Regulation and Governance.
67. Wang G. (2020). Humans in the Loop: The Design of Interactive AI Systems.
68. Weiner M. (2019). The Conscience of Artificial Intelligence. Porsche Engineering Magazine,
Issue 2/2019 https://newsroom.porsche.com/en/2020/technology/porsche-engineering-
artificial-intelligence-professor-christoph-luetge-19462.html.
69. White Paper on Artificial Intelligence: A European approach to excellence and trust. (2020).
European Commission.
70. Whittlestone J, Nyrup R, Alexandrova A, Dihal K., & Cave S. (2019). Ethical and societal
implications of algorithms, data, and artificial intelligence: a roadmap for research’, Nuffield
Foundation.
71. Winner L. (1980). Do Artifacts Have Politics?. Daedalus 109(1), 121–136.
72. Zittrain JL. (2006). The Generative Internet. Harvard Law Journal, 119.

18

Electronic copy available at: https://ssrn.com/abstract=3609292