Ethics for AI in Business

Saïd Business School, University of Oxford:

Felipe Thomaz ([email protected]), Natalia Efremova, Francesca Mazzi,
Greg Clark, Ewan MacDonald, Rhonda Hadi, Jason Bell and Andrew Stephen

Report commissioned by the ICC Research Foundation.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Ethics in AI for Business

Introduction

At this point in time, it would be a severe understatement to simply point out that the

recent rapid development and adoption of artificial intelligence (AI) processes has been

transformative in all facets of business operations. For a variety of reasons, from the digital

transformation of businesses and the accompanying greater data availability to the acceleration

in computing power; the variety of methodologies and approaches that are broadly classified

as AI have drastically shifted the way in which firms ingest information, process and analyse

data, and augment or automate tasks and decisions. In fact, more broadly, the advent of AI is

likely to have a substantial impact on the nature of consumer behaviour, firm competition, and

future successful business models (Thomaz et al 2020).

While customers and firms have benefited from these technologies, the overall value of

AI has been significantly larger. Beyond purely economic, cost, and efficiency advantages, the

application of these technologies has brought vast social benefits, including advances in new

product development that take into account improved accessibility for individuals with

disabilities, and reduced waste. AI has been applied in healthcare for both the development of

new treatments and the improvement of diagnostics, and amongst others in the development of

a vaccine to contrast the COVID-19 pandemic. In natural sciences, it has been used to analyse

the pattern of natural fires or improving irrigation in farming. It has even been used in

conservation for the identification of individual animals and their habitats.

However, we also have seen a number of unintended consequences and negative

outcomes, with numerous reports of innovative attempts by companies and governments falling

short in spectacular and public fashion. Among these are chatbots who learned hate speech

online, human resources processes that unwittingly enshrined sexual discrimination in code, or

school testing scores being assigned on basis of socioeconomic condition rather than student

Electronic copy available at: https://ssrn.com/abstract=3871867

achievement. Invariably, this has led to new policy debates on how to curb negative outcomes

and the concept of ethical use of AI.

Given the tremendous incentives for businesses to develop and implement AI systems,

the private sector is a critical player in these policy debates. And consequentially, any delay or

inability for the private sector in adopting ethical use of AI further motivates stronger

regulation of the technology and its uses, which in turn would have an expected dampening

effect on further innovation and advancement in AI. Unsurprisingly, many companies are

tackling this issue head on; especially those whose business models are more strongly

intertwined with the development and use of AI processes. These firms have taken the lead in

ethical AI thought and leadership, setting up governance structures, statements of purpose, and

guidance materials for their own organizations and peers. Similarly, IGOs and governments

have added their own voices and guidance on the ethical application of AI as it pertains to the

citizens or a combined vision of what an Ethical AI society should aspire to be.

However, while firms, governments, intergovernmental and non-governmental

organizations have developed and published materials on ethical AI principles, a series of

significant issues have become apparent. In a review of 84 existing ethical AI guidelines, Jobin

and colleagues (2019) identify that there remains a significant divergence in understanding and

interpretation of the principles being discussed. Simply, there is no common language and

meaning across many of these documents, no agreement on the issues, domains, or actors to be

considered. Furthermore, this analysis highlights a sense of confusion due to unintegrated

sectorial guidelines and strong geographic bias, with the bulk of guidelines representing North

American and European perspectives, rather than a global framework. And lastly, but perhaps

most importantly, none of the existing work presents a viable candidate application strategy on

how to achieve and manage AI ethics.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Here we aim to resolve these issues. This effort motivated the partnership between the

International Chamber of Commerce (ICC) and the University of Oxford’s Saïd Business

School to collaborate and provide the voice of business and actively promote the ethical use of

AI. We start from the foundational belief that ethical design of AI is intrinsically good for

business, and that its application is ultimately good for broader society. Businesses of all sizes,

sectors and locations must be empowered to seize the many new opportunities that AI

technologies offer. This report provides the necessary guidance to help avoid the pitfalls

associated with these technologies and ensure the effective implementation of ethical AI for

the benefit of the business, its stakeholders and the wider community. We especially

recommend this report, to the attention of small and medium enterprises (SMEs) thinking of

applying AI to their business or as a thought starter for companies of all sizes just setting out

on their AI journey.

In order to summarize and organize the existing guidance on Ethical AI in a practical

framework we first collected any items that were published in academic journals on the topic,

but also considered business statements (e.g. Microsoft Responsible AI Principles, Google AI

Principles and Responsibilities), governmental guidance documents (e.g. G20 AI Principles),

as well as those prepared by non-governmental and intra-governmental agencies (e.g. OECD

AI Principles and AI Policy Observatory, A Framework for Ethical AI at the United Nations),

and even articles found in the popular press. For those interested in specifics, the literature in

AI ethics in computer science from 1954 to 2021 is described and reviewed in the Annex 1 of

this document. This review identified eight key areas of concern when discussing AI ethics,

from Privacy to Algorithmic fairness and beyond. These are highlighted and described more

completely in Annex 2 of this document. Annex 3 summarizes the key insights from this

review.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Lastly, in order to fully take advantage of the collective knowledge and the wealth of

information contained in all of these documents, we turned to AI ourselves and used Natural

Language Processing techniques to process, organize, and extract relational information from

the documents available (the corpus). Details on text normalization, frequency analysis, vector

embeddings, named entity recognition and knowledge graph generation, as well as and some

results, are given in Annex 4 of this document.

But importantly, by scouring the existing knowledge base on Ethical AI and combining

the analysis of this corpus with operating ethical frameworks, we were able to organize the

critical components into a defensible and implementable workflow. We discuss this next.

Our AI Implementation Roadmap

1. Set of principles to be implemented throughout the lifecycle of AI

2. Definition of the principles and their hierarchy

3. Assessing ethical risks (general)

4. Practical Implementation

- Mission Statement

- Ethics design plan

A. Phase and application specific ethical impact assessment

B. Mitigation strategies

C. Monitoring system

1. Set of principles to be implemented throughout the lifecycle of AI

There are a number of issues that become apparent when we consider the field of ethical

AI and its application to business, especially when we must translate discussions into actual

tangible steps and processes in our operations. First and foremost is the nature of ethical

principles themselves. These are invariably intangible, grand ideas with rather fuzzy

boundaries, which makes them perhaps easy to grasp in general, but difficult to pin down with

Electronic copy available at: https://ssrn.com/abstract=3871867

any sort of precision. For example, consider the idea of “justice.” We might have an intuitive

sense of what being “just” might be, or we might know it when we see it, but the specifics of

what makes something “just” becomes more difficult.

Relatedly, there is ample confusion among similar terms and concepts. For example,

do you need to provide “justice” or “fairness”? Are they the same thing? What about Ethical

AI, versus Responsible AI, or Trustworthy AI? These terms are often used interchangeably,

depending on the speaker, the document, or domain. Sometimes, the same word might be used

for an ethical principle and the reverse meaning in computer science applications (e.g.

“transparency”). We might then be particularly worried that different individuals or

organisations might read the same statement; but leave with entirely different interpretations

of its meaning.

Additional confusion about enabling, creating, and maintaining ethical AI applications

could be attributed to the significant, and perhaps overwhelming, number of ethical principles

that are used to explain ethical considerations and design. In the aforementioned overview of

ethical guidelines for AI (Jobin, Ienca and Vayena 2019), the authors’ analysis of 84 guidance

documents identified 1180 distinct ethical concepts, belonging to 13 ethical themes and

categories. As such, the ethical designer must contend with a host of difficulties to define

concepts.

Lastly, and further complicating the issue, these concepts are always presented in an

unintegrated, non-relational form. Meaning that while we might have guidance with lists of

concepts to consider, we do ultimately lack any frameworks with which to guide the design of

ethical AI business applications and governance systems. By analogy, a firm’s directory

contains all of the people that make up the organization, but hardly captures the tasks, teams

and structure that allows the business to operate successfully. And similarly, a directory of

Electronic copy available at: https://ssrn.com/abstract=3871867

ethical principles can be useful, but perhaps insufficient for adaptation and inclusion into

specific and diverse business contexts.

2. Definition of the principles and their hierarchy

In order to develop our framework and organization of ethical principles, we conducted

a review of ten descriptive and normative ethical theories (Utilitarianism, Deontology, Virtue

Ethics, Ethics of Care, Egoism, Divine Command, Natural Law, Social Contract, Theory of

Justice, and Moral Relativism) in order to understand their arguments and structures.

Additionally, we reviewed and took inspiration from the Belmont report, a guiding document

on Protection of Human Subjects in Biomedical and Behavioural Research. This report

represents a longstanding gold standard in the academic community for the design of processes

where individuals will have less power, or agency in their environments. Lastly, we derived

insights from our own Natural Language Processing analysis of AI ethics documents, policy

discussions, guidelines, and academic publications as described earlier. Importantly, this last

analysis allows us to better select the most relevant ethical principles for our context, and to

begin organizing them in a coherent structure.

Figure 1 below presents the resulting framework, and more precise definitions of each

component follows.

Importantly, we suggest that these ethical principles are hierarchical; meaning, that

many are components and subcomponents of other larger and more complex principles. We

begin at the highest level of the hierarchy with “Ethical.” The second layer of the framework,

then provides the means by which one would achieve this highest-order ethical behavior.

Specifically, it suggests that ethics requires one to be both “Responsible” and “Accountable.”

More precise definitions follow below, but for now, responsibility relates to the execution of

one’s duties and responsibilities faithfully. Accountability; however, reflects one’s ability and

Electronic copy available at: https://ssrn.com/abstract=3871867

willingness to explain and justify their actions and decisions. Therefore, while responsibility

speaks to the means by which ends are achieved, accountability is concerned with the ends

themselves (intended, perceived, and actual). Ethical AI applications in business then require

both a responsible (how) and accountable (what outcomes) approach.

Why might each of these components be insufficient on their own, when in fact, a

number of organizations globally have adopted a “responsible AI” approach and language?

Our approach looks at two domains: being accountable with respect to end results, and being

responsible, with respect to the execution of your duties and overall behaviour. Conversely,

simply executing duties and responsibilities faithfully, is necessary, but insufficient. Attention

must be paid to the outcomes of those actions. Our framework suggests that Ethical AI in

business should satisfy both conditions: responsible means, and accountable outcomes. Taken

together, the two concepts aim at safeguarding not only that organisations ensure the proper

functioning of the AI systems that they design, develop, operate or deploy throughout their

entire lifecycle, in accordance with their roles and applicable regulatory frameworks, but also

that they demonstrate this commitment through their actions, behaviour and decision-making

process more broadly.

The next layer in our hierarchy of principles proves to be the most important, and one

in which we’ll focus most of our attention. Which principles are required in order to have a

“Responsible” application? These are: “Human-Centric,” “Fair” and “Harmless.” You’ll note

that these principles are also quite vague, which suggests that each will in turn still be made up

of several smaller but increasingly more tangible ideas. And precisely, the next set of principles

in our hierarchy provide the mean by which one achieves these goals. For example, “Human

Centric” is achieved via transparent, intelligible, and sustainable systems, but also includes to

concept of beneficence. “Fair” processes are those that can be classified as just, inclusive, and

non-discriminatory. In turn, “Harmless” systems are safe, robust, and private. Importantly,

Electronic copy available at: https://ssrn.com/abstract=3871867

these lower-level characteristics are not only more easily understood and defined, but also more

measurable, meaning that at this level we might be able to start introducing controls and

assurances of performance within acceptable bounds. For example, we might easily be able to

quantify the safety of a system based on its failure rate, as well as its robustness based on the

scope of consequences given a failure. Those values, together with an assessment of the

systems privacy risk, aggregate upwards toward a defensible classification of a “Harmless” AI

process. If we can equally support the idea that the process is “Fair” and “Human-Centric”,

then we can aggregate from there and summarize the process as “Responsible.” And so on until

an overarching stance of it being an ethical AI application.

“Accountable” deserves special attention as it featured heavily in our NLP analysis of

the body of work discussing AI ethics in business. While the concept appears prominently in

legal and regulatory texts, it is featured a lot less in in business-based publications. That should

not be altogether surprising, as regulators will invariably focus on the need for new regulation.

That is their function after all. Our focus was to unpack the reasons for the term’s relative

unpopularity in business literature. It is our belief that this disconnect comes from the close

association or perhaps even confusion of this concept with that strictly of liability and

culpability. In that sense, to be held accountable would be a discussion on being legally liable

for the consequence of AI processes and certain outcomes. And while that is certainly one

component of accountability, which is made more complicated by regional legal differences

and buyer/supplier of AI solution relationships; accountability is more complex and more

important than this suggests.

Our framework places a number of principles and activities as components of

accountability: proactive leadership, reporting, contesting, correcting, and liability. Focusing

on this first component, proactive leadership, we recognize that accountability is a leadership

function and not strictly reactive where the concern lies in addressing culpability when

Electronic copy available at: https://ssrn.com/abstract=3871867

something goes wrong. Leadership, and specifically with the business and new AI development

space, is tied to a forward-looking view, scenario mapping, as well as an awareness of

consequences such that the firm is accountable to its various stakeholders, internal and external.

In that sense, accountability is not strictly a legal concern, but rather an obligation to one’s

customers, employees, communities, as well as the public at large.

Lastly, we note that the framework itself and the list of items aggregating up to the main

and critical concepts of “human centric,” “fair,” and “harmless” is open-ended. As noted

earlier, with over one thousand candidate ideas, a comprehensive and exhaustive framework

would be both unwieldy and perhaps unwanted. While additional ideas and ethical concepts

exist, they will invariably fit as components of one of these three main general principles, or

perhaps as a component of accountability itself. This organization then presents a reduction

and a simplified view of the field of AI ethics for business applications, allowing us to build

practices and governance structures to ensure ethical processes as well as outcomes.

Figure 1: Hierarchy of Principles

Electronic copy available at: https://ssrn.com/abstract=3871867

Principles – definitions

In this section we provide brief definitions and explanations of the principles presented in our

framework:

Ethical - moral principles that govern a person's behaviour or the conducting of an activity.

Responsible - having an obligation to do something, or having control over or care for

someone, as part of one's job or role.

Accountable - required or expected to justify one’s intentions, actions or decisions.

Human-Centric - respects individuals’ value and agency, including any necessary.

information for self-determination, and care for the environment and conditions where

individuals exist.

Transparent - Open to necessary scrutiny.

Intelligible - Able to be understood; comprehensible.

Sustainable - Able to operate in ecological balance, avoids depleting resources.

Beneficence – Accomplishing a good outcome, increasing the welfare of the user.

Fair - an ability to judge without reference to one's feelings or interests, treating people

equally without favouritism or discrimination.

Just - deserved or appropriate treatment in the circumstances, well founded opinions and

appraisals; justifiable.

Inclusive - not excluding any groups of people.

Non-Discriminatory - no prejudicial distinctions between different categories of people.

Harmless – Not able, or likely to cause harm.

Safe - operation/use does not cause harm, protected from risk.

Robust - The process or application is said to be robust if it fails “gracefully” rather than

“spectacularly.” Usually it means the system is not sensitive to extreme examples, outliers,

Electronic copy available at: https://ssrn.com/abstract=3871867

and adversarial attacks, combined with fail-safe measures that allow it to fail (if it does)

without causing harm.

Private - provides the ability to control access to information about the self

3. Assessing Ethical Risks

Three risk “buckets”: Data, Algorithm, Business Use

(1) Data - The term “ethical AI” implies a commitment to verify and ensure that the

entire workflow of AI respects ethical principles. Indeed, such term shall be seen as a label, a

guarantee of adequacy that the business is willing to achieve and maintain, and that

consequently contributes to build trust between users and firms that utilises AI. The required

commitment shall cover all the relevant phases and all the aspects surrounding the use of AI,

through a so-called “ethical impact assessment”.

For the purpose of designing a meaningful impact assessment, a first macro-

categorization of areas that need to be considered when assessing ethical risks includes data,

algorithm and business use.

The first category, data, represents an essential component of AI that provides

substance to the use of AI in terms of content. From the training of the AI to the actual

deployment of it, data plays the role of nourishing the AI with the relevant information that

will initially provide the AI with actual knowledge and reference, and then contribute in a

determinant way to the final outcomes of its deployment.

To provide an example, the AI can learn how to differentiate apples from oranges based

on provided data, which can include pictures of oranges and apples and related labels to identify

it. Moreover, when the AI is used to practically differentiate apples from oranges, it will be

able to do so in relation to images or pieces of information of oranges and apples that it is

provided with.

Electronic copy available at: https://ssrn.com/abstract=3871867

 As a consequence, the selection of data that the machine is provided with (so-called

data set) represents a fundamental aspect of the AI functioning, but also a potential source of

unethical outcomes. For example, if the AI is trained only on a certain type of apples, it will

not be able to recognise other types of apples, which will result in a discrimination. If the AI is

used only in relation to these two types of fruit, it should not be deployed in environments or

for purposes that might include other kind of food.

Hence, the first identified macro-area of ethical risk, data, requires a substantial analysis

of the selection of information provided to the AI. Such analysis shall be performed in light of

the general purpose of the use of AI within the business, the environment of the actual

deployment, and the ethical principles identified, and it shall concern all data provided to the

AI from the training phase to the actual use.

Moreover, ethical concerns that specifically arise in relation to the category of data

relate to data protection law and the right to privacy. With the proliferation of digitalised

information, the legislation regarding data protection flourished in many countries in the past

years. Laws and regulations on data protection shall be carefully assessed and considered not

only in relation to the immediately applicable legislation, for example the legislation of the

country where the firm is legally based, but also in relation to the geographical scope of

operation of the business. Indeed, for example, legislation such as the General Data Protection

Regulation is also applicable to businesses based outside the EU that process personal data of

EU citizens.

Consequently, an adequate ethical assessment of AI requires an evaluation of the type

of data that is processed by AI in order to evaluate whether it involves information that

concerns individuals, whether the individuals are identifiable through the processed

information, whether the information concerns sensitive aspects of individuals such as health

or religion. Moreover, such assessment will need to consider the requirements of the applicable

Electronic copy available at: https://ssrn.com/abstract=3871867

regulations, the potential risks for individuals, if any, from a privacy perspective, and the

required measures to prevent risks and ensure an effective exercise of data subjects rights.

(2) Algorithms - The second category, the algorithm, represents the basic component

of AI. The algorithm is, indeed, a set of instructions, and the AI is composed of algorithms that

can be changed and adapted based on the intelligent learning of the AI. Hence, algorithms

represent the bone structure of the AI. As such, they do represent a potential source of unethical

output as well. The entirety of the lifecycle of algorithms should be monitored to verify the

respect of ethical principles.

Indeed, for example, biases in the AI can result from the phase of development of the

algorithm. Biases can arise due to unethical coding influenced by pre-existent biases, coming

from social institutions, cultural practices, and personal attitudes, perpetuated by the

programmer. Additionally, during the development phase, technical biases can be incorporated

in the design of the algorithm. Moreover, during its entire lifecycle, the AI can be biased due

to emergent biases, since as it was mentioned earlier the algorithms that compose the AI can

be changed and adapted through the learning of AI. Hence, if the AI learns an unethical pattern,

which can be caused by the real-world use or by processing biased data, it might adapt its

algorithms based on such unethical information, therefore incorporating, and perpetuating such

biases. Although certain AI systems, such as neural networks, contain a high number of layers

and algorithms cannot be monitored closely during their deployment due to the so-called black-

box effect, in such cases appropriate measures should be in place to constantly monitor the

ethical deployment of such models.

Consequently, an adequate ethical impact assessment requires a close monitoring of the

different phases of the algorithms, from their development to their deployment, to the extent

that biases can be detected and corrected.

Electronic copy available at: https://ssrn.com/abstract=3871867

 (3) Business Use - The third category, business use, comprises the purpose and the use

of the AI that its owner intends to pursue, and the actual deployment in the real world.

The ethical principles set out in relation to AI should be in principle applicable to the

business as well. In order to use the “ethical AI” label as described above, adjectives such as

“non-harmful” and “fair” should not only be referable to the AI within a business, but also to

the business model employing the AI per se. For example, an AI could be well-trained to avoid

gender biases, but it will not be ethical if used to monitor employees in working conditions that

violate human rights.

In this sense, the business use of AI can be a source of unethical behaviours for two

main reasons. The first, immediate reason concerns the business goal, as described in the

example. If the purpose of the development of AI is the creation of a lethal weapon, the

outcome of the AI is foreseeably unethical.

The second reason relates to the actual deployment of the AI in the real world. The AI

can be built to achieve ethical goals, but it can prove to produce unethical outputs when opened

to the public if risks are not adequately prevented. As an example, discriminations or harms

can derive from the interaction of users with the AI or between themselves through the AI. A

mismatch of values and intentions between the function that the AI is intended to perform and

its actual deployment in the real world can result in unethical output and consequently lead to

liability issues.

Hence, in order to prevent liability issues and to properly address potential risks

deriving from the business use of the AI, a careful consideration of the societal consequences

of the use of AI is fundamental. An adequate ethical impact assessment requires an evaluation

of the objectives and goals to be achieved with AI, and of the risks associated with the real-

world use of the AI to achieve such a purpose.

Electronic copy available at: https://ssrn.com/abstract=3871867

4. Practical Implementation

While potentially interesting as a thought exercise alone, we propose that the

framework for ethical principles presented in this work is much more impactful in practice if

used for the design of an ethical plan for the firm’s AI ambitions and for each of its specific

applications. Specifically, this approach would require two steps:

First, the creation of a general, external and internal facing statement or certification of

the firm’s stance as an adherent of AI ethics and guaranteeing the integrity of its practices. In

terms of scope and tone, this document might be seen an analogous to a firm’s purpose or

mission/vision statement. In fact, ethical AI policies should be a logical extension of, and

entirely consistent with, corporate purpose. This initial statement should serve as blanket

declaration that the firm’s AI technology is consistent as such, and ethically applied. For the

purpose of enhancing transparency, the statement should include the hierarchy of values of the

firm, that are used to address trade-offs.

As a second step, we create a set of supporting documents that provide validity to the

general statement. Each is an internal application-specific ethical design plan, where the

designer carefully audits the AI process, raising questions and flagging risks as well as

concerns around data, algorithm, and business use. Once specific ethical concerns and risks are

identified, specific controls as well as management and mitigation strategies for each risk can

be introduced. As part of the supporting documents, the firm should keep record of decisions

concerning trade-offs, that should reflect the values indicated in the mission statement, to

facilitate external auditing and enhance transparency. The exercise provides the opportunity to

carefully consider the sources of ethical tension, depending on the application and context, and

to create a manageable control system that increases the likelihood of human centric, fair, and

harmless outcomes; therefore, responsible and more easily accountable processes.

Electronic copy available at: https://ssrn.com/abstract=3871867

 For the firm, the first document is a “flight plan” by analogy. A guiding clearance that

all is in order in terms of trajectory and the underlying engineering. Keeping to the same

analogy, the second set of documents provides us our “flight checklist”, or a more precise

listing of important safety checks and considerations pre-flight along with processes for in-

flight monitoring. While the first document, "the flight plan", incorporates the mission

statement as a stable long-term trajectory, the second set of documents is dynamic in nature as

it evolves with the use of AI. By monitoring the use of AI, new tensions or risks can be

identified, and the ethical "checklist" can be updated accordingly with related safeguards and

preventive measures. In this way, the second set of documents works in loop cycles that ensures

constant monitoring and updating of the ethical plan.

As a governance process, this approach provides a number of benefits. Broad adoption

would legitimize the practice and severely minimize the threat of regulatory force, who would

otherwise need to step in to consolidate practices and protect end users. A well-studied, thought

out, and implemented ethical design should minimize the risk exposure of the firm by

identifying, minimizing, and mitigating issues before they have an opportunity to become

unpleasant and large issues. And lastly, as AI becomes as increasingly strong competitive

battleground for businesses, the evidence of good ethical practices should serve as a

competitive advantage over firms that choose to forego such measures.

As such, the firm would have the following document workflow:

A) A general AI Ethics Statement for the whole Organization (external facing)

B) An application specific control document for each use-case (internal), identifying

a. Ethical risk identification and impact assessment

b. Mitigation strategies for identified concerns

c. Monitoring system for the application that account for identified risks.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Any changes to operations would then trigger a quick check on whether new risks are

introduced, or potential impact is altered, against the noted components of the framework given

in Figure 1. Anything that threatens the execution of Ethical AI then requires some work on

mitigation strategies and monitoring updates in order to continue operating in a way that

benefits the firm and its innovation, but also keeping with its responsibility and accountability

to stakeholders. And altogether, this should move AI applications towards a much more

familiar management practice around precise controls and margins of acceptable operation,

rather than vague aspirational requirements.

References

1. Thomaz, F., Salge, C., Karahanna, E. et al. Learning from the Dark Web:
leveraging conversational agents in the era of hyper-privacy to enhance
marketing. J. of the Acad. Mark. Sci. 48, 43–63 (2020).
https://doi.org/10.1007/s11747-019-00704
2. Jobin, Anna, Ienca, Marcello, and Vayena, Effy (2019), 'The global landscape
of AI ethics guidelines', Nature Machine Intelligence, 1 (9), 389-99.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Annex 1 – Literature Review

This section describes literature on ethics, fairness and transparency in AI and related

fields, such as machine learning (ML) and data science (DS). In order to provide a full

description of literature on this topic(s), we include the following items: (1) an overview of the

volumes of the publications over the period of time, covered with electronic publishing; (2) a

comparison of the literature from computer science versus other fields; and (3) a classification

and a brief description of the main topics in computer science literature.

In order to scope a body of literature on all three topics of interest, we conducted a scoping

review of the existing corpus of documents on AI ethics. This included a search for literature

containing principles and guidelines for ethical AI from academic sources. We analyse non-

academic sources of literature with knowledge graphs in the second part of this research (see

NLP model for knowledge representation with graphs).

A scoping review is a method aimed at synthesizing and mapping the existing literature

(Arksey, H. and O’Malley 2005) that is considered particularly suitable for complex or

heterogeneous areas of research (Pham et al. 2014).

First, we performed a keyword-based search on Scopus using the following keywords:

‘AI ethics’, ‘AI fairness’ and ‘AI transparency’. First search returned 858 results; second search

returned 156 and the third search returned 316 results (Table A1.1). The overlapping results

constituted only 7 studies, which indicates that these keywords are used in different scientific

domains. The results below confirm this assumption.

Additionally, we analysed unique source titles (journal or conference titles). In the first group,

we have found 478 unique source titles, in the second group we have found 112 sources and in

the third one 249.

Electronic copy available at: https://ssrn.com/abstract=3871867

Table A1.1 Number of papers per keyword group: AI ethics, AI fairness and AI transparency.

We analyse the total number of publications and additionally we distinguish between computer

science literature versus other types of publications.

Keywords Results Source titles Years CS papers non-CS papers

AI ethics 858 478 1954 - 2021 280 578

AI fairness 156 112 1999 - 2020 117 39

AI transparency 316 246 1983 - 2020 176 140

For every category, we have checked whether the source was from computer science

literature or not: we performed binary labelling on the retrieved data. We marked as technical

literature only the sources, where the scope of the publisher was explicitly technical, and

marked as non-technical literature everything else1. The results showed significant growth of

publications for all fields over time. To demonstrate the growth in the number of publications

per year, we have displayed the following data below:

- number of sources per year (growth) per keyword pair;

- number of technical sources (computer science, CS) per year (growth) per keyword pair.

To provide a clearer picture of the growth of the volumes of publications, we have

combined them into four groups: (1) everything that was published before 2000; (2) 2000-

2009; (3) 2010- 2018 and (4) 2019-2020.

1 For example, Ethics and Information Technology journal was marked as non-technical
(https://www.springer.com/journal/10676 ).

Electronic copy available at: https://ssrn.com/abstract=3871867

Table A1.2. Publications per year group. Total: total number of publications per year; CS -

publications in computer science per year.

a. AI ethics

b. AI fairness

c. AI

transparency

From tables A1.1 and A1.2, we can see that AI ethics discussion shifted significantly

from computer science to adjacent fields. Before 1999, the related literature in computer

science provides over 99% of the whole corpus, whereas over the last 2 years 40% of published

studies moved to other domains, including medicine, sociology, etc.. A similar picture can be

Electronic copy available at: https://ssrn.com/abstract=3871867

observed for non-computer science literature on the “AI Transparency” category, showing a

shift from 0.02% to 38% of studies. However, for AI Fairness, the percentage of non-computer

science papers is still quite low, representing only 13% of publications.

In order to fully describe the literature on AI ethics in computer science literature, we

reviewed multiple sources, including workshops on the high-impact computer science

conferences (NeurIPS, ICML, ICLR, CVPR, IEEE, AAAI and IJCAI), popular online courses

(on AI and Data Science Ethics) and all available journal titles. Then, after consultation with

researches in ethics and computer science, we constructed the following list of the main

research topics, related to AI ethics:

1. Data Ownership

2. Privacy

3. Anonymity

4. Data Validity

5. Algorithmic Fairness

6. Model Transparency

7. Platform Responsibility

8. Data Licensing

Electronic copy available at: https://ssrn.com/abstract=3871867

 Annex 2 – Key areas of AI Ethics

1. Data Ownership. The first topic that we can distinguish in the field of AI ethics, fairness and

transparency is data ownership.

The terms “ownership” and “property rights” in the context of data science ethics are

defined as following: “property rights (as I will deal with them here, are the rights of

ownership” (Becker 1980). Property refers to the thing(s) to which property rights apply

(Hummel et al 2020). In broader context, it covers questions on whom the data belongs to

(data ownership), how to store it and how to use it correctly (aggregation, ownership

management and privacy protection). Literature on data ownership covers such questions as

direct data ownership; data aggregation operations in data analytics, machine learning and

artificial intelligence (Rosa et al 2020); dynamic ownership management and privacy

protection (Janeček 2018), (Yuan et al. 2018).

2. Privacy. This topic overlaps with the previous one (data ownership). However this category

is more related to privacy-preserving algorithms and secure processing of data (Ryffel et al,

2018, Adams. 2014, Boyd 2011, Davidson 2015, Walter 2002). This field covers GDPR,

privacy policies and regulations in different fields in different countries, therefore it has a very

large volume of documents.

3. Anonymity. This field covers such topics and de- and re-identification and related issues

(Witten, 2011). De-identification is the removal of identifiable information from data. Re-

identification is the process by which anonymized personal data is matched with its true owner.

In order to protect the privacy interests of consumers, personal identifiers, such as name and

social security number, are often removed from databases containing sensitive information.

This field also includes studies on face recognition. In this context, re-identification problem

Electronic copy available at: https://ssrn.com/abstract=3871867

describes matching faces across multiple sources for the purpose of sequential authentication

over space and time (Knyaz 2019, Wechsler and Li 2014).

4. Data Validity. Data validity covers appropriate use of data science methods, specifically

from the field of statistics in real-world application, for example for medical records and

adjacent fields. Additional sub-items within Validity can also include data accuracy, relevance

and timeliness to action/need, as well as its completeness.

5. Algorithmic Fairness. This field covers identifying, measuring and improving algorithmic

fairness when using AI algorithms and heavily concentrates on the algorithmic bias (including

bias in images, texts, spatial data etc.) that can lead to multiple problems, such as unfairness in

decision making and racial or class discrimination (Pessach and Shmueli 2020).

6. Model transparency. This field covers issues of machine learning models, related to both

model transparency and data transparency (Ananny and Crawford 2018). The data transparency

problem is related to the fact that end users of a model are often unaware of what data was used

to train the model. Since the majority of ML models require vast amounts of data, this

encourages situations where data is collected and combined from across various and unrelated

sources. This might lead to class imbalance, which affects the resulting performance of the

model. And while regulation and policies on data gathering might be seen to solve this problem

in the long term; there is currently no way with which we can ensure that the model producer

uses the reliable data for training ML models. Another issue is the transparency of the

algorithm itself. A model, especially a deep learning model, is very difficult to decode once

it’s been trained. In many cases, there is a possibility to look inside the model and visualise

what is happening inside deep learning model, giving rise to a whole separate field of computer

Electronic copy available at: https://ssrn.com/abstract=3871867

vision who works solely on this problem (Olah 2015, Erhan et al. 2008, Simonyan et al. 2013,

Zeiler and Fergus 2014).

7. Platform responsibility. This field regards platforms as moral agents, covering translating

of the basic moral principles; analysis of the literature from different fields, such as philosophy

and sociology in application to artificial agents, including robots, ML models, AI models and

machines in general. A separate subfield deals with the representation of social media platform

as moral agents and discusses the responsibilities that platforms should carry versus acting as

mediators in human communication (Awad 2018, Floridi and Sanders 2004, Moor 2006).

8. Data licensing. This field covers issues of licensing of data in the fields of artificial

intelligence and machine learning. There is no common framework for data licensing akin to

the licensing of open source software. Two main issues can be distinguished: lack of data

transparency and conceptual ambiguities in existing licensing language. Solving these issues

may help foster fairer and more efficient markets for data through bringing about clearer tools

and concepts that better define how data can be used in the fields of AI and ML (Benjamin

2019).

Electronic copy available at: https://ssrn.com/abstract=3871867

 Annex 3 – Key insights from AI ethics documents

One immediate and clear issue that we can see is the fairly obvious division in

understanding, and ultimately in priorities, when it comes to AI, the underlying technology, its

applications, and its consequences. This finding confirms the conclusions of the narrower

review done by Jobin and colleagues (2019), and as such, it is not entirely surprising. However,

in this analysis we are better able to identify the nature and themes of this divergence. First,

there is the issue of time. Business discussion appears to de-emphasize the impact of real time

applications, focusing instead on decision support. The distinction is important, as it goes to

the heart of the discussion around automation, and specifically the proportion of system

suggestions that are immediately accepted by managers (thus shrinking the difference between

decision support and automation). Real-time applications encompass customer-facing

applications where data, systems, and processes are encoded and enshrined inside of AI

umbrella, and leaves little room for course correction or adjustment by humans. Therefore, it

should be clear why regulators might be more interested in these situations, and more likely to

take regulatory actions against such activity.

The second crucial distinction lies in an overall underrepresentation of “accountability”

as a topic in the business discourse around AI, and its narrow use in the legal sense of the word,

or as “liability” equivalent. This is understandable from a number of perspectives, including

the outdated legal schemes to address liability that prove to be inefficient when applied to black

box AI. However, as we discuss in the main document, accountability is also associated with a

leadership component that can serve as a base for proactive engagement in ethical AI during a

transition phase.

By closing these gaps in understanding and refocusing the discussion, business will be

better able to assume the leadership not only in the development of AI, but also in discussions

on its governance.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Annex 4 – Analysis of Ethical AI Documentation

NLP Analysis of Ethics in AI documentation

Since the literature on AI ethics is significant, we used a machine learning approach to

retrieve relevant information from the documents and to understands similarities and

differences between those texts.

We processed 2 text corpora: business and regulatory literature corpus (divided by

groups of stakeholders) and news corpus (divided by industry sectors).

Figure 1 Text corpora used in NLP analysis consists of "Business" publications and "News" corpus.

In order to extract meaningful data from the large text corpus, we perform two types of

processing: information retrieval (IR) and knowledge graph (KG) creation. Text processing

was divided into 3 main parts: (1) pre-processing of the text to make it readable for ML

algorithms, (2) IR and (3) KG creation. These techniques were performed on each text corpus

and the results were compared between the industry sectors in case of news corpus and groups

of stakeholders in case of business and regulatory literature corpus.

Electronic copy available at: https://ssrn.com/abstract=3871867

 First, we normalized texts and created a dictionary of words from all texts to perform

frequency analysis. It allowed us to understand what words (unigrams2) and pairs (bigrams) or

triplets (trigrams) of words are used the most, thus uncovering the most important topics from

the text (word-level text processing). Next, we analysed unigrams and bigrams in the text,

extract text frequencies and created word clouds. In the third part, we extracted information

from text to create knowledge graphs. In this task, we work on the sentence level to extract

main relationships from the text. These relationships were then further used to build knowledge

graphs (Jurafsky and Martin 2008).

For information retrieval, we first performed a series of operations, collectively called

text normalisation. We begin this process with cleaning all texts with regular expressions:

removing special symbols, hashtags, emoticons, etc., since they are not important for this

research. We then separate our texts into words (or tokenize words from running text, the task

of tokenization). English words are often separated from each other by whitespace, but

whitespace is not sufficient in this case: we also separate punctuation (pad punctuation symbols

with whitespaces etc.), expand clitic contractions that are marked by apostrophes (using Penn

Treebank tokenization standard). Another part of text normalization is lemmatization, the task

of determining that two words have the same root, despite their surface differences. For

example, the words sang, sung, and sings are forms of the verb sing. The word sing is the

common lemma of these words, and a lemmatizer maps from all of these to sing. Stemming

refers to a simpler version of lemmatization in which we mainly remove suffixes from the end

2 An n-gram is a sequence of N n-gram words: a unigram is a separate word (‘intelligence”); a 2-gram (or bigram)
is a two-word sequence of words like “artificial intelligence”, ”generalised”, and a 3-gram (or trigram) is a three-
word sequence of words like “generalised artificial intelligence”. We use bigrams and unigrams for extracting
frequencies from text.

Electronic copy available at: https://ssrn.com/abstract=3871867

of the word. In this case we use lemmatization. Next, we performed information retrieval3 to

extract frequent words and phrases from texts. Finally, we used sentence segmentation to break

up our texts into individual sentences, using cues: question marks, periods and exclamation

points.

In the description below, we will use language models (models that assign probabilities

to sequences of words, LMs). We first removed the stop words from our list of tokens, created

in the previous step. Stop words are very frequent words like “the” and “a”. We use frequency–

inverse document frequency (TFIDF) vectorizer for frequency analysis of bigrams and

unigrams. TFIDF, is a numerical statistic that is intended to reflect how important a word is to

a document in a collection or corpus. The TFIDF value increases proportionally to the number

of times a word appears in the document and is offset by the number of documents in the corpus

that contain the word, which helps to adjust for the fact that some words appear more frequently

in general. We use this method to count all the occurrences of each word in the text corpus (for

bigram analysis, we do the same for bigrams) and analyse the results. Results are displayed in

fig. 2 through 5 below.

3 Information retrieval (IR) is the task of finding the document d from the D documents in some collection that
best matches a query q. For IR we’ll therefore also represent a query by a vector, also of length |V|, and we’ll need
a way to compare two vectors to find how similar they are.
This process of information extraction (IE), turns the unstructured information embedded in texts into structured
data, for example for populating a relational database to enable further processing.

Electronic copy available at: https://ssrn.com/abstract=3871867

Figure 2 Unigram analysis on the text corpus (business literature)

Figure 3 Bigram analysis on the text corpus (business literature)

Electronic copy available at: https://ssrn.com/abstract=3871867

 Figure 4 Unigram analysis on the text corpus (news)

Figure 5 Bigram analysis on the text corpus (news)

Knowledge graphs are used to visually represent a text or a collection of texts, based

on main topics that are discussed in the text. To build a knowledge graph, we first find all

names or named entities in a text. The task of named entity recognition (NER) is to find each

Electronic copy available at: https://ssrn.com/abstract=3871867

mention of a named entity in the text and label its type. We use long-short term memory

(LSTM) (Hochreiter and Schmidhuber, 1997) models to extract following entities: people,

places, and organizations. Next, we extract relations: find and classify semantic relations

among the text entities. These are often binary relations like “child-of”, “employment”, part-

whole, and geospatial relations. For simplicity, we work on a sentence level and extract only

the basic relations from the text (source, target and relation triplets).

We construct a data frame, that contains extracted entities and build knowledge graphs

based on the extracted entities. Sample graph is depicted in Fig. 6. Nodes represent named

entities and connections represent relations between the nodes. For our goal, relations are

unimportant, since we want to understand what is discussed, not to extract the sentiment.

Figure 6 Knowledge graph, extracted from a subset of texts.

Since the whole graph is difficult to read, we extract only the relations, where the source

is ‘AI’ (Fig.7). We do the same for each group of texts in each text corpus. Resulting graphs

allow us to quickly visualise connections between the most important concepts in the text.

Electronic copy available at: https://ssrn.com/abstract=3871867

 Figure 7 Knowledge graph depicting the relationships of the term 'AI' in the

“News – Healthcare” corpus.

The proposed method allowed us to process the large corpus of texts and define

differences and similarities between different groups of texts in an automatic manner. This

approach allows not only to deal with large text corpus, but seamlessly add new data and repeat

the analysis as frequently as required.

Electronic copy available at: https://ssrn.com/abstract=3871867

References

1. Arksey, Hilary and O'Malley, Lisa (2005), 'Scoping studies: towards a

methodological framework', International Journal of Social Research
Methodology, 8 (1), 19-32.

2. Pham, Mai T., et al. (2014), 'A scoping review of scoping reviews: advancing
the approach and enhancing the consistency', Research Synthesis Methods, 5
(4), 371-85.

3. List of conferences: https://neurips.cc; https://icml.cc; https://iclr.cc;

http://cvpr2021.thecvf.com; https://www.ieee.org/conferences/index.html;
https://ijcai-21.org ; https://aaai.org/Conferences/AAAI-21/

4. Becker, Lawrence C (1980), 'The moral basis of property rights', NOMOS:

Am. Soc'y Pol. Legal Phil., 22, 187.

5. Becker, Lawrence C (1980), 'The moral basis of property rights', NOMOS:

Am. Soc'y Pol. Legal Phil., 22, 187.

6. Hummel, Patrik, Braun, Matthias, and Dabrock, Peter (2020), 'Own Data?
Ethical Reflections on Data Ownership', Philosophy & Technology.

7. Rosa, Marco, Cerbo, Francesco Di, and Lozoya, Rocio Cabrera (2020),
'Declarative Access Control for Aggregations of Multiple Ownership Data',
Proceedings of the 25th ACM Symposium on Access Control Models and
Technologies (Barcelona, Spain: Association for Computing Machinery), 59–
70.

8. Janeček, Václav (2018), 'Ownership of personal data in the Internet of

Things', Computer Law & Security Review, 34 (5), 1039-52.

9. Yuan, Haoran, et al. (2018), 'DedupDUM: Secure and scalable data

deduplication with dynamic user management', Information Sciences, 456,
159-73.

10. Ryffel, Theo, et al. (2018), 'A generic framework for privacy preserving deep
learning', arXiv preprint arXiv:1811.04017.

11. Adams, A. 2014. Report of a debate on Snowden's actions by ACM members.

SIGCAS Computers & Society 44, 3: 5-7.

12. Boyd D., Crawford K. 2011. Six provocations for big data. Retrieved from
http://papers.ssrn.com/sol3/Papers.cfm?abstract_id=1926431

13. Davidson, Roei and Poor, Nathaniel. 2015. The barriers facing artists use of
crowdfunding platforms: Personality, emotional labor, and going to the well
one too many times. New Media & Society 17, 2: 289-307.
https://journals.sagepub.com/doi/abs/10.1177/1461444814558916

Electronic copy available at: https://ssrn.com/abstract=3871867

14. Joseph Walther. 2002. Research ethics in internet-enabled research: Human
subjects issues and methodological myopia. Ethics and Information
Technology 4, 3: 205-216.
https://link.springer.com/article/10.1023/A:1021368426115

15. Ian H. Witten, Mark A. Hall. What’s It All About?, in Data Mining (Third
Edition), 2011

16. Knyaz, VA, Maksimov, AA, and Novikov, MM (2019), “Vision Based
Automated Anthropological Measurements and Analysis”, International
Archives of the Photogrammetry, Remote Sensing & Spatial Information
Sciences.

17. Wechsler, Robert T, et al. (2014), 'Conversion to lacosamide monotherapy in

the treatment of focal epilepsy: results from a historical‐controlled,
multicenter, double‐blind study', Epilepsia, 55 (7), 1088-98.

18. Pessach, Dana and Shmueli, Erez (2020), “Algorithmic fairness”, arXiv
preprint arXiv:2001.09784.

19. Ananny, Mike and Crawford, Kate (2018), “Seeing without knowing:
Limitations of the transparency ideal and its application to algorithmic
accountability”, new media & society, 20 (3), 973-89.

20. Olah, Christopher (2015), “Understanding LSTM networks”.

https://colah.github.io/posts/2015-08-Understanding-LSTMs/

21. Erhan, D., Bengio, Y., Courville, A. and Vincent, P., 2009. Visualizing
higher-layer features of a deep network University of Montreal, Vol 1341,
pp. 3.

22. Simonyan, Karen, Vedaldi, Andrea, and Zisserman, Andrew (2013), 'Deep
inside convolutional networks: Visualising image classification models and
saliency maps', arXiv preprint arXiv:1312.6034.

23. Zeiler, Matthew D and Fergus, Rob (2014), 'Visualizing and understanding
convolutional networks', European conference on computer vision (Springer),
818-33.

24. Awad, Edmond, et al. (2018), 'The moral machine experiment', Nature, 563
(7729), 59-64.

25. Floridi, Luciano and Sanders, Jeff W (2004), 'On the morality of artificial
agents', Minds and machines, 14 (3), 349-79.

26. Moor, James (2006), 'The Nature, Importance, and Difficulty of Machine
Ethics', IEEE Intelligent Systems, 21, 18-21.

27. Benjamin, Ruha (2019), 'Race after technology: Abolitionist tools for the new
jim code', Social Forces.

Electronic copy available at: https://ssrn.com/abstract=3871867

28. Jobin, Anna, Ienca, Marcello, and Vayena, Effy (2019), 'The global landscape
of AI ethics guidelines', Nature Machine Intelligence, 1 (9), 389-99.

29. Jurafsky, Daniel and Martin, James H (2008), 'Speech and Language
Processing: An introduction to speech recognition, computational linguistics
and natural language processing', Upper Saddle River, NJ: Prentice Hall.

30. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory.
Neural Comput. 9, 8 (November 15, 1997), 1735–1780.
DOI:https://doi.org/10.1162/neco.1997.9.8.1735

Electronic copy available at: https://ssrn.com/abstract=3871867