The Underlying Values of Data Ethics Frameworks: A Critical Analysis of Discourses and Power Structures

1 Introduction

During the last decade, numerous headlines and (news) articles about discriminating algorithm decision-making, harmful applications or data breaches have shown the drawbacks of technological progress. Those incidents have raised awareness for the impact of algorithmic bias and set ethical questions on the public agenda (Fast and Horvitz 2016). Since 2016, a growing number of reports deals with the subjects of artificial intelligence (AI) and ethics (Whittlestone et al. 2019b). Around the same time, many private and public organizations addressed the concern towards harmful technologies by releasing codes of ethics and guidelines. Those declarations are non-enforceable soft law codes and usually voluntary commitments or general recommendations (Haas and Gießler 2020). The increasing number of codes and guidelines manifests the organizations’ interest in shaping ethical guidance according to their preferences (Jobin, Ienca, and Vayena 2019) but also leads to accusations of “ethics washing” (Floridi 2019).

However, the codes of ethics proved to be rather abstract and not helpful in guiding ethical decision making in daily work tasks (Morley et al. 2020; O’Boyle 2002). Lately, there have been a few publications that are explicitly designed for practical application (e.g. Center for Democracy & Technology 2017; DataEthics.eu 2021; Department for Digital, Culture, Media & Sport UK 2018; Schäfer and Franzke 2020; Tarrant and Maddison 2021). In this paper, they will be referred to as Data Ethics Frameworks (DEF) and are meant to address ethical issues during the early stages of development and design of a technological system or service. The DEFs are targeted not only at engineers and developers but to the persons who are involved in a specific project which makes use of data or applies algorithmic systems. The frameworks’ appealing design and their brief questionnaire genre facilitate the ethical deliberation process and are the major differences compared to conventional guidelines and codes.

The purpose of the current study is to identify the values and value conflicts conveyed via the DEFs. Applying the Critical Discourse Analysis (CDA) according to Fairclough (1995) to four frameworks allows to further investigate how the values and value conflicts are presented and which discursive patterns are used. In this respect, DEFs are considered an expression of the social practices in the Computer and Information Ethics (CIE) discipline and thus are shaped by the corresponding beliefs and processes. Consequently, the relations and power structures between the involved actors and stakeholders should be critically examined. Finally, this holistic approach enables discussing the frameworks’ contribution to social change. In a matter that is developing so rapidly and changing our society so profoundly, it is of particular interest to know what motives prevail and who is taking part in shaping the discourse.

The originality of the study is obtained by the qualitative approach and the chosen methodology since ethical guidelines have been analyzed predominantly quantitatively with content analysis (Hagendorff 2020; Jobin, Ienca, and Vayena 2019; Schiff et al. 2021) or with frame analysis (Greene, Hoffman, and Stark 2019). Considering value conflicts as well as practices and attitudes of the social field achieve an amplified understanding of issues at stake. Power asymmetries were discussed on a broad level in previous studies (Jobin, Ienca, and Vayena 2019; Schiff et al. 2020). Thanks to a methodology that is sensible towards ideological language and hegemonic views from the perspective of oppressed groups, new insights on power structures are added.

This paper first gives a brief overview of the foundations of CIE and currently discussed subjects. The second part introduces the CDA methodology and the proceeding of the four analyzed frameworks. Subsequently, the identified values, value conflicts, and discourses are reported, following a discussion on the underlying structures and assumptions. Finally, limitations of the study and recommendations for future research are given.

2 Literature Review

Debates on ethical challenges arise especially in times of significant technical progress. Beginning with early developments of computers, Wiener stated concerns regarding automated work forces and application in warfare in the 1940s and 1950s (Bynum 2010). In the 1960s, Parker and Maner called attention to the increasing crimes committed with the help of computers and by technicists (Bynum 2010). Moor (1985) provides an explanation for the numerous ethical problems caused by computers. The author characterizes computers as logically malleable which appears limitless in the potential application. For activities that have not been able without computers people cannot rely on good practices or ethical standards, a state of “policy vacuum” as Moor (1985) calls it. The author sees the role of computer ethics in dissolving the conceptual muddle by ethical reflection of concrete cases and providing a “coherent conceptual framework within which to formulate a policy for action” (Moor 1985, 266). Since 1985, the level of abstraction has shifted from technological means to information as content to data as the smallest entity. The focus followed the technical development and drew the attention to the points where ethical problems are likely to arise (Floridi and Taddeo 2016, 3). Still, Moor’s understanding of the mandate remains valid nowadays. It can be traced in the definition of Data Ethics proposed by Floridi and Taddeo (2016) which determines it as “the branch of ethics that studies and evaluates moral problems related to data […], algorithms […] and corresponding practices […], in order to formulate and support morally good solutions (e.g. right conducts or right values)” (3). There seems to be deficient consensus regarding the denomination as the term “Data Ethics” is one among others – including the authors of the same definition making use of expressions like “AI ethics” in more recent articles (e.g. Morley et al. 2020).

Albeit the terminology remains ambiguous, Moor’s understanding of the discipline’s contribution serves to classify the existing research. The existing research on the one hand studies ethical challenges and on the other hand fathoms suggestions for right conduct. Ethical deliberation in CIE is often drawing on normative ethics, a branch “that is concerned with establishing how things should or ought to be, how to value them, which things are good or bad, and which actions are right or wrong” (Dignum 2017, 3). Virtue ethics, deontology, and consequentialism are the normative theories that are most prominently referred to, focusing on the morality of the acting person, the act itself or the outcome (Dignum 2017; Kraemer, van Overveld, and Peterson 2011; Mittelstadt et al. 2016; Saltz and Dewar 2019; Sandvig et al. 2016). A loose interpretation of all three normative theories turns out most promising in practice for guiding ethical assessment in CIE, as Ananny (2016) and Sandvig et al. (2016) recommend.

3 Methodology

The acknowledgment of the reciprocal influence of technology and values indicates a constructivist notion. Consequently, a constructivist procedure, namely Critical Discourse Analysis (CDA), is employed to examine four Data Ethics Frameworks. The specificities of the qualitative method will be presented as well as sampling of the frameworks, analysis of language patterns, and creation and publication of texts.

Social constructivism can be described as the idea “that our knowledge of the world, including our understanding of human beings, is a product of human thought rather than grounded in an observable, external reality” (Burr 2015, 222). As persons we are formed by the culture, norms, and situations surrounding us. Language as a key aspect of social interaction is therefore essential to social constructionism and results in a number of theories and approaches centering on discourse (Burr 2015). One of those is CDA by Fairclough (1995), a well-established concept that supports studying power structures in language and takes an interactional perspective towards meaningful social change.

CDA approaches are generally characterized by their assumption of ideological language, the constructive relation between language and social practice, as well as the critical view from the perspective of the oppressed group (Jørgensen and Phillips 2002). CDA as coined by Fairclough builds theoretically on Marxist academics for his understanding of ideology and hegemony. Ideology not only is perceivable in language but also in social practices and becomes invisible through the recognition as common-sense (Fairclough 1995). The competition of the prevailing social class surfaces at the level of discursive practice i.e., the activities in which language and texts are embedded. This so-called hegemonic struggle “contributes in varying degrees to the reproduction or transformation of the existing order of discourse, and through that of existing social and power relations” (Fairclough 1995, 77). Thus, people have a range of possibilities to act, to use language creatively, change meanings, and to resist.

Fairclough’s CDA framework considers three dimensions of discourse corresponding to different discourse analysis techniques (Figure 1). Text is at the core as it is embedded into certain practices of text production, dissemination, and interpretation (discourse practice) and constructed by customs, beliefs, and conduct in a specific social field (sociocultural practice). To detect the power relations and ideological strains in language, a linguistic analysis is conducted at text level. Discourse practice is addressed by a critical review of the established processes of text production. The inclusion of social theory permits deducing explanations of the consequences of reproduction or change for the wider sociocultural practice (Jørgensen and Phillips 2002).

Figure 1:

Fairclough’s CDA framework and analysis methods (in black) applied to the current study (in blue). Own representation based on Fairclough (1995, 98).

4 Results

In this section the identified values and value conflicts are reported and illustrated with the help of brief examples. Furthermore, the prevailing discourses and power structures are assessed and complemented with discursive practices of the organizations.

In total, 646 codes were assigned to the four frameworks. Of this sum, 194 codes were distributed to values and value conflicts. Evidence for all values of the integrated list was found except for environmental welfare. However, dignity, trustworthiness, and solidarity were not referred to under those denominations which made them difficult to distinguish from other values and thus were merged. On the contrary, transparency was singled out as a separate value for its prominent appearance. Various discourses had been detected due to corresponding vocabulary: a business, legal, technological, anti-discrimination, democratization, and general value discourse (246 codes). The remaining 233 codes were assigned to the “designers” – the target group of the frameworks – and the “stakeholders” who are mentioned as objects to technologies and devices. Power structures manifest in the presentation of this group of persons. Table 2 gives insights on the quantitative distributions of the codes among DEFs.

5 Extracted Values and Discourses

Human welfare is often referred to as “benefit” and “user need” which should guide the development process and is present in three of four frameworks. Present tense indicates certainty of the additional value (“What are the benefits of the project?”, DEDA, l. 12). Emphasis is placed on the communication and enhancement of positive impact which follows a business logic (“How are you measuring and communicating positive impact? How could you increase it?”, DEC, l. 40–41). Benefit is therefore juxtaposed with business advantage and financial revenue. In discursive practice, this corresponds with the positive storytelling communication especially applied by the Open Data Institute (2019b).

In contrast to human welfare, non-maleficence is determined more specifically since causes of harm are supposedly more familiar. Bias, misuse, and misinterpretation are prominently mentioned but use of modal verbs and passive forms intend to create a distance to the project and disguise responsible actors (“What are the problems or concerns that might arise in connection with this project?”, DEDA, l. 13). The notions of misuse and misinterpretation imply the ambition to uphold authority of “right” use and interpretation. Harmful incidents seem to be a potential business risk also in terms of public criticism (“Does the project risk generating public concern or outrage?”, DEDA, l. 43). Non-maleficence also includes knowing one’s limits and consulting external experts. That aspect is well covered in DEW although expertise appears to be closely related with formal education (“subject matter experts”, DEW, l. 36). Finally, a precaution principle is observed among all frameworks to consider possible long-term implications.

Justice is regarded in terms of fair and equal treatment and freedom of discrimination and thereby underpins dignity. DEDA refers to justice and inclusion as values, while other frameworks mention eradication of bias at the level of data, algorithm, and outcomes (“Where could bias have come into this analysis?”, DD Tool, l. 92). The broad coverage of bias mirrors the extensive public and academic debate and relates to further activities by the organizations. The Center for Democracy and Technology for instance carried out a project in that domain whose results directly inspired DD Tool (Lange 2016) and Utrecht Data School employ a new project under the acronym BIAS (Utrecht Data School 2021a). The questions are on the one side reinforcing a “bad actor frame” (Hoffman 2019, 903) by locating the source of bias in the assumptions of single persons or homogeneous teams and on the other side transport a tendency of technological determinism that humans are surrendered to (Greene, Hoffman, and Stark 2019). Solutions for achieving justice are framed by a technology discourse as optimizing technologies are considered (“Did your feedback mechanism capture and report anomalous results in a way that allows you to check for biased outcomes?”, DD Tool, l. 84). An anti-discrimination discourse is perceptible when people are given the opportunity to share their experience and are taken seriously (“Do citizens have the opportunity to raise objections to the results of the project?”, DEDA, l. 45).

La Fors, Custers, and Keymolen (2019) see transparency as backing accountability but the frameworks recognize transparency separately. It is understood in terms of publishing openly and communicating understandably (“Could you publish your methodology, metadata, datasets, code or impact measurements?”, DEC, l. 60). Therefore, it is related to a business discourse making use of strategic communication (“What is the communication strategy with regard to this project?”, DEDA, l. 40). A legal discourse comes into play as certain duties are exemplified (“Are non-deterministic outcomes acceptable given your legal or ethical obligations around transparency and explainability?”, DD Tool, l. 52). Organizations themselves handle openness differently: Utrecht Data School and Central Digital & Data Office publish their frameworks and supplementary material freely accessible (newest versions) whereas some reports of Center for Democracy & Technology needed to be retrieved from an internet archive. Open Data Institute, the one that emphasized transparency and openness, offers a free download of the framework, but the user guide and other publications are accessible from a commercial platform only after registration. Furthermore, organizations are often reluctant at openly disclosing their motives, contributors, and understanding of Data Ethics.

Accountability is used interchangeably with responsibility and is aimed at ensuring traceability. This results in a distribution of responsibility for ethical challenges to individuals (“Is there a person on your team tasked specifically with identifying and resolving bias and discrimination issues?”, DD Tool, l. 9) or among organizational hierarchies (“How often will you report on these plans to senior reporting officers?”, DEW, l. 58). A considerable legal discourse demonstrates that responsibility is often interpreted with regards to existing liabilities (“Which laws and regulations apply to your project?”, DEDA, l. 35). Relation with transparency indicates documentation or preparation for audition. The organizations that released the frameworks refuse to take any accountability for outcomes and implications of the ethical deliberation (Broad, Smith, and Wells 2017; Utrecht Data School 2020b).

Autonomy, the ability to pursuing own thoughts, will, goals, and decision-making, starts at the very beginning of being involved into a certain project or data collection (“Was the data collected in an environment where data subjects had meaningful choices?”, DD Tool, l. 28). Furthermore, it is about the means of interaction within the project and with the creators of a technology or device. Untargeted collaboration is mentioned in two frameworks (“Are you routinely building in thoughts, ideas and considerations of people affected in your project? How?”, DEC, l. 65). In many phrases, stakeholders are referred to either pertaining to a passive collective or according to their sensitive features. While the feelings and thoughts of the designers are given room, stakeholders are not deemed the same position to utter feelings apart from discrimination. In practice of framework creation, the organizations headed their development, but acquired support from the practice (Utrecht Data School 2020a) or at least conducted user studies for revision (Central Digital & Data Office 2020; Ginnis et al. 2016). Caring about instruction and adjusting it to different groups strengthens democratization since knowing an issue supports forming of opinions and acting autonomously (“What information or training might be needed to help people understand data issues?”, DEC, l. 66).

Privacy is understood as sensitive personal information and predominantly addressed in the frame of a legal discourse (“If using personal data, do you understand obligations under data protection legislation?”, DEW, l. 13). The European General Data Protection Regulation (GDPR) is strict in that sense to protect individuals and work towards data minimization. Additionally, it is referred to other elements of GDPR and shows the influence of regulations even in an ethical context that could go further than the requirements of law (“Have you conducted a PIA (Privacy Impact assessment [sic!]) or DPIA (Data Protection Impact Assessment)?”, DEDA, l. 49). Since present tense indicates the normalization of processing of personal data, data minimization is an aspect that entails an alternative (“How can you meet the project aim using the minimum personal data possible?”, DEW, l. 18). Moreover, anonymization techniques, access control mechanisms, and synthetic data are listed as options and illustrate the technical discourse applied to that value.

DEDA is the only framework that encourages reflecting on personal and organizational values and considers that it may give inconsistencies and conflicts. This issue was adjusted due to experiences in practice (Franzke, Muis, and Schäfer 2021). Other frameworks vaguely refer to conflicts of interest between project aims (“Are you replacing another product or service as a result of this project?”, DEC, l. 33) and stakeholder groups (“Is there a fair balance between the rights of individuals and the interests of the community?”, DEW, l. 24). However, little guidance is given to examine how “fair” might be interpreted and how the various interests are documented. Values may come into conflict where supposed project benefits interfere with individuals’ privacy or autonomy or among the project team (“Are all parties involved in agreement as to this strategy?”, DEDA, l. 40).

Overall, these results show that the identified values of human welfare, non-maleficence, justice, transparency, accountability, autonomy, and privacy are addressed in all frameworks. The principles are often listed singularly which strengthens the impression that value conflicts are omitted. Surprisingly, the general value discourse is marginal compared to the business, technology, legal, anti-discrimination, and democratization discourse. At the level of discursive practice, the issuing organizations headed framework creation although various stakeholders were often included at later stages of the process. Particularly in terms of transparency, many organizations do not comply with their own ambitions transported in the DEFs. Apparently, due to the focus on the perspective of their target group, anyone who deals with data, it is out of sight that stakeholders and data subjects may support certain values as well.

6 Discussion

The findings of this analysis show similarities with other studies that examined ethical guidelines (Hagendorff 2020; Jobin, Ienca, and Vayena 2019; Schiff et al. 2021). Values like transparency, privacy, accountability, and justice are predominantly mentioned across the four frameworks and likewise are the most covered ones in the mentioned studies. On the contrary, values are not opposed among each other nor are conflicts between groups or interests amplified. The academic debates in CIE are reflected by the same importance of topics as principles are better researched than value conflicts. This connection demonstrates the mutual construction of discourse and sociocultural practice. In the tradition of social constructionism and CDA, implications for knowledge creation and power structures are discussed in the following.

7 Conclusions

In this study, four practically designed Data Ethics Frameworks from public or non-profit institutions were investigated to identify the promoted values and evaluate the representation of value conflicts. Findings show a set of established values which is present in all frameworks although emphasis differs across the publications. This indicates a close relation with the information ethics discipline which has been occupied from the beginning with preserving human values. Although language structures and values indicate a reinforcement of established practices and customs, a hegemonic struggle between various actors can be observed. Values are increasingly interpreted as a business factor and thus related to aspects of communication, legal compliance, and technological solutions. Concerns on eradicating discrimination allow for an anti-discrimination and democratization discourse but reinforced power asymmetries weaken the effectiveness. Since the frameworks take the perspective of their target group, affected data subjects are contemplated from a distance and not meaningfully included into the debate, neither via DEFs at text level nor in terms of discursive practices in the moment of text creation. Therefore, it is recommended to apply and test means of participation of diverse direct and indirect stakeholders because it is still unexplored in research. Focusing on intersection of values and value conflicts should also play a greater role in research. This has the potential to advance the academic and public debate in the question of which values should be prioritized, and which trade-offs might be acceptable at designing future technologies.