The Ethics of AI Ethics: An Evaluation of Guidelines

Research in the field of AI ethics ranges from reflections on how ethical principles can be implemented in decision routines of autonomous machines (Anderson and Anderson 2015; Etzioni and Etzioni 2017; Yu et al. 2018) over meta-studies about AI ethics (Vakkuri and Abrahamsson 2018; Prates et al. 2018; Boddington 2017; Greene et al. 2019; Goldsmith and Burton 2017) or the empirical analysis on how trolley problems are solved (Awad et al. 2018) to reflections on specific problems (Eckersley 2018) and comprehensive AI guidelines (The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems 2019). This paper mainly deals with the latter issue. The list of ethics guidelines considered in this article therefore includes compilations that cover the field of AI ethics as comprehensively as possible. To the best of my knowledge, a few preprints and papers are currently available, which also deal with the comparison of different ethical guidelines (Zeng et al. 2018; Fjeld et al. 2019; Jobin et al. 2019). While especially the paper from Jobin et al. (2019) is a systematic scoping review of all the existing literature on AI ethics, this paper does not aim at a full analysis of every available soft-law or non-legal norm document on AI, algorithm, robot, or data ethics, but rather a semi-systematic overview of issues and normative stances in the field, demonstrating how the details of AI ethics relate to a bigger picture.

The selection and compilation of 22 major ethical guidelines were based on a literature analysis. This selection was undertaken in two phases. In the first phase, I searched different databases, namely Google, Google Scholar, Web of Science, ACM Digital Library, arXiv, and SSRN for hits or articles on “AI ethics”, “artificial intelligence ethics”, “AI principles”, “artificial intelligence principles”, “AI guidelines”, and “artificial intelligence guidelines, following every link in the first 25 search results, while at the same time ignoring duplicates in the search process. During the analysis of the search results, I also sifted through the references in order to manually find further relevant guidelines. Furthermore, I used Algorithm Watch’s AI Ethics Guidelines Global Inventory, a crowdsourced, comprehensive list of ethics guidelines, to check whether I missed relevant guidelines. Via the list, I found three further guidelines that meet the criteria for the selection. In this context, a shortcoming one has to consider is that my selection is biased towards documents which are western/northern in nature, excluding guidelines which are not written in English.

I rejected all documents older than 5 years in order to only take guidelines into account that are relatively new. Documents that only refer to a national context—such as for instance position papers of national interest groups (Smart Dubai Smart Dubai 2018), the report of the British House of Lords (Bakewell et al. 2018), or the Nordic engineers’ stand on Artificial Intelligence and Ethics (Podgaiska and Shklovski)—were excluded from the compilation. Nevertheless, I included the European Commission’s “Ethics Guidelines for Trustworthy AI” (Pekka et al. 2018), the Obama administration’s “Report on the Future of Artificial Intelligence” (Holdren et al. 2016), and the “Beijing AI Principles” (Beijing Academy of Artificial Intelligence 2019), which are backed by the Chinese Ministry of Science and Technology. I have included these three guidelines because they represent the three largest AI “superpowers”. Furthermore, I included the “OECD Principles on AI” (Organisation for Economic Co-operation and Development 2019) due to their supranational character. Scientific papers or texts that fall into the category of AI ethics but focus on one or more specific aspects of the topic were not considered either. The same applies to guidelines or toolkits, which are not specifically about AI but rather about big data, algorithms or robotics (Anderson et al. 2018; Anderson and Anderson 2011). I further excluded corporate policies, with the exception of the “Information Technology Industry AI Policy Principles” (2017), the principles of the “Partnership on AI” (2018), the IEEE first and second version of the document on “Ethically Aligned Design” (The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems 2016, 2019), as well as the brief principle lists of Google (2018), Microsoft (2019), DeepMind (DeepMind), OpenAI (2018), and IBM (Cutler et al. 2018) which have become well-known through media coverage. Other large companies such as Facebook or Twitter have not yet published any systematic AI guidelines, but only isolated statements of good conduct. Paula Boddington’s book on ethical guidelines (2017) funded by the Future of Life Institute was also not considered as it merely repeats the Asilomar principles (2017).

The decisive factor for the selection of ethics guidelines was not the depth of detail of the individual document, but the discernible intention of a comprehensive mapping and categorization of normative claims with regard to the field of AI ethics. In Table 1, I only inserted green markers if the corresponding issues were explicitly discussed in one or more paragraphs. Isolated mentions without further explanations were not considered, unless the analyzed guideline is so short that it consists entirely of brief mentions altogether.

As shown in Table 1, several issues are unsurprisingly recurring across various guidelines. Especially the aspects of accountability, privacy or fairness appear altogether in about 80% of all guidelines and seem to provide the minimal requirements for building and using an “ethically sound” AI system. What is striking here is the fact that the most frequently mentioned aspects are those for which technical fixes can be or have already been developed. Enormous technical efforts are undertaken to meet ethical targets in the fields of accountability and explainable AI (Mittelstadt et al. 2019), fairness and discrimination aware data mining (Gebru et al. 2018), as well as privacy (Baron and Musolesi 2017). Many of those endeavors are unified under the FAT ML or XAI community (Veale and Binns 2017; Selbst et al. 2018). Several tech-companies already offer tools for bias mitigation and fairness in machine learning. In this context, Google, Microsoft and Facebook have issued the “AI Fairness 360” tool kit, the “What-If Tool”, “Facets”, “fairlern.py” and “Fairness Flow”, respectively (Whittaker et al. 2018).

Accountability, explainability, privacy, justice, but also other values such as robustness or safety are most easily operationalized mathematically and thus tend to be implemented in terms of technical solutions. With reference to the findings of psychologist Carol Gilligan, one could argue at this point that the way AI ethics is performed and structured constitutes a typical instantiation of a male-dominated justice ethics (Gilligan 1982). In the 1980s, Gilligan demonstrated in empirical studies that women do not, as men typically do, address moral problems primarily through a “calculating”, “rational”, “logic-oriented” ethics of justice, but rather interpret them within a wider framework of an “empathic”, “emotion-oriented” ethics of care. In fact, no different from other parts of AI research, the discourse on AI ethics is also primarily shaped by men. My analysis of the distribution of female and male authors of the guidelines, as far as authors were indicated in the documents, showed that the proportion of women was 41.7%. This ratio appears to be close to balance. However, it should be considered that the ratio of female to male authors is reduced to a less balanced 31.3% if the four AI Now Reports are discarded, which come from an organization that is deliberately led by women. The proportion of women is lowest at 7.7% in the FAT ML community’s guidelines which are focused predominantly on technical solutions (Diakopoulos et al.). Accordingly, the “male way” of thinking about ethical problems is reflected in almost all ethical guidelines by way of mentioning aspects such as accountability, privacy or fairness. In contrast, almost no guideline talks about AI in contexts of care, nurture, help, welfare, social responsibility or ecological networks. In AI ethics, technical artefacts are primarily seen as isolated entities that can be optimized by experts so as to find technical solutions for technical problems. What is often lacking is a consideration of the wider contexts and the comprehensive relationship networks in which technical systems are embedded. In accordance with that, it turns out that precisely the reports of AI Now (Crawford et al. 2016, 2019; Whittaker et al. 2018; Campolo et al. 2017), an organization primarily led by women, do not conceive AI applications in isolation, but within a larger network of social and ecological dependencies and relationships (Crawford and Joler 2018), corresponding most closely with the ideas and tenets of an ethics of care (Held 2013).

What are further insights from my analysis of the ethics guidelines, as summarized in Table 1? On the one hand, it is noticeable that guidelines from industrial contexts name on average 9.1 distinctly separated ethical aspects, whereas the average for ethics codes from science is 10.8. The principles of Microsoft’s AI ethics are the most brief and minimalistic (Microsoft Corporation 2019). The OpenAI Charta names only four points and is thus situated at the bottom of the list (OpenAI 2018). Conversely, the IEEE guideline contains the largest volume with more than 100.000 words (The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems 2019). Finally, yet importantly, it is noteworthy that almost all guidelines suggest that technical solutions exist for many of the problems described. Nevertheless, there are only two guidelines which contain genuinely technical explanations at all—albeit only very sparsely. The authors of the guideline on the “Malicious Use of AI” provide the most extensive commentary here (Brundage et al. 2018).

Despite the fact that the guidelines contain various parallels and several recurring topics, what are issues the guidelines do not discuss at all or only very occasionally? Here, I want to give a (non-exhaustive) overview of issues that are missing. Two things should be considered in this context. First, the sampling method used to select the AI ethics guidelines has an effect on the list of issues and omissions. When deliberately excluding for instance robot ethics guidelines, this has the effect that the list of entries lacks issues that are connected with robotics. Second, not all omissions can be treated equally. There are omissions which are missing or severely underrepresented without any good reason—for instance the aspect of political abuse or “hidden” social and ecological costs of AI systems—, and omissions that can be justified—for instance deliberations on artificial general intelligence or machine consciousness, since those technologies are purely speculative.

Nevertheless, in view of the fact that significant parts of the AI community see the emergence of artificial general intelligence as well as associated dangers for humanity or existential threats as a likely scenario (Müller and Bostrom 2016; Bostrom 2014; Tegmark 2017; Omohundro 2014), one could argue that those topics could be discussed in ethics guidelines under the umbrella of potential prohibitions to pursue certain research strands in this area (Hagendorff 2019). The fact that artificial general intelligence is not discussed in the guidelines may be due to the fact that most of the guidelines are not written by research groups from philosophy or other speculative disciplines, but by researchers with a background directly in computer science or its application. In this context, it is noteworthy that the fear of the emergence of superintelligence is more frequently expressed by people who lack technical experience in the field of AI—one just has to think of people like Stephen Hawking, Elon Musk or Bill Gates—while “real” experts generally regard the idea of a strong AI as rather absurd (Calo 2017, 26). Perhaps the same holds true for the question of machine consciousness and the ethical problems associated with it (Lyons 2018), as this topic is also omitted from all examined ethical guidelines. What is also striking is the fact that only the Montréal Declaration for Responsible Development of Artificial Intelligence (2018) as well as the AI Now 2019 Report (2019) explicitly addresses the aspect of democratic control, governance and political deliberation of AI systems. The mentioned documents are also the only guidelines that explicitly prohibits imposing certain lifestyles or concepts of “good living” on people by AI systems, as it is for example demonstrated in the Chinese scoring system (Engelmann et al. 2019). The former document further criticizes the application of AI systems for the reduction of social cohesion, for example by isolating people in echo chambers (Flaxman et al. 2016). In addition, hardly any guideline discusses the possibility for political abuse of AI systems in the context of automated propaganda, bots, fake news, deepfakes, micro targeting, election fraud, and the like. What is also largely absent from most guidelines is the issue of a lack in diversity within the AI community. This lack of diversity is prevailing in the field of artificial intelligence research and development, as well as in the workplace cultures shaping the technology industry. In the end, a relatively small group of predominantly white men determines how AI systems are designed, for what purposes they are optimized, what is attempted to realize technically, etc. The famous AI startup “nnaisense” run by Jürgen Schmidhuber, which aims at generating an artificial general intelligence, to name just one example, employs only two women—one scientist and one office manager—in its team, but 21 men. Another matter, which is not covered at all or only very rarely mentioned in the guidelines, are aspects of robot ethics. As mentioned in the methods chapter, specific guidelines for robot ethics exist, most prominently represented by Asimov’s three laws of robotics (Asimov 2004), but those guidelines were intentionally excluded from the analysis. Nonetheless, advances in AI research contribute, for instance, to increasingly anthropomorphized technical devices. The ethical question that arises in this context echoes Immanuel Kant’s “brutalization argument” and states that the abuse of anthropomorphized agents—as, for example, is the case with language assistants (Brahnam 2006)—also promotes the likelihood of violent actions between people (Darling 2016). Apart from that, the examined ethics guidelines pay little attention to the rather popular trolley problems (Awad et al. 2018) and their alleged relation to ethical questions surrounding self-driving cars or other autonomous vehicles. In connection to this, no guideline deals in detail with the obvious question where systems of algorithmic decision making are superior or inferior, respectively, to human decision routines. And finally, virtually no guideline deals with the “hidden” social and ecological costs of AI systems. At several points in the guidelines, the importance of AI systems for approaching a sustainable society is emphasized (Rolnick et al. 2019). However, it is omitted—with the exception of the AI Now 2019 Report (2019)—that producer and consumer practices in the context of AI technologies may in themselves contradict sustainability goals. Issues such as lithium mining, e-waste, the one-way use of rare earth minerals, energy consumption, low-wage “clickworkers” creating labels for data sets or doing content moderation are of relevance here (Crawford and Joler 2018; Irani 2016; Veglis 2014; Fang 2019; Casilli 2017). Although “clickwork” is a necessary prerequisite for the application of methods of supervised machine learning, it is associated with numerous social problems (Silberman et al. 2018; Irani 2015; Graham et al. 2017), such as low wages, work conditions and psychological work consequences, which tend to be ignored by the AI community. Finally, yet importantly, not a single guideline raises the issue of public–private partnerships and industry-funded research in the field of AI. Despite the massive lack of transparency regarding the allocation of research funds, it is no secret that large parts of university AI research are financed by corporate partners. In light of this, it remains questionable to what extent the ideal of freedom of research can be upheld—or whether there will be a gradual “buyout” of research institutes.

The close link between business and science is not only revealed by the fact that all of the major AI conferences are sponsored by industry partners. The link between business and science is also well illustrated by the AI Index 2018 (Shoham et al. 2018). Statistics show that, for example, the number of corporate-affiliated AI papers has grown significantly in recent years. Furthermore, there is a huge growth in the number of active AI startups, each supported by huge amounts of annual funding from Venture Capital firms. Tens of thousands of AI-related patents are registered each year. Different industries are incorporating AI applications in a broad variety of fields, ranging from manufacturing, supply-chain management, and service development, to marketing and risk assessment. All in all, the global AI market comprises more than 7 billion dollars (Wiggers 2019).

A critical look at this global AI market and the use of AI systems in the economy and other social systems sheds light primarily on unwanted side effects of the use of AI, as well as on directly malevolent contexts of use. These occur in various areas (Pistono and Yampolskiy 2016; Amodei et al. 2017). Leading, of course, is the military use of AI in cyber warfare or regarding weaponized unmanned vehicles or drones (Ernest and Carroll 2016; Anderson and Waxman 2013). According to media reports, the US government alone intends to invest two billion dollars in military AI projects over the next 5 years (Fryer-Biggs 2018). Moreover, governments can use AI applications for automated propaganda and disinformation campaigns (Lazer et al. 2018), social control (Engelmann et al. 2019), surveillance (Helbing 2019), face recognition or sentiment analysis (Introna and Wood 2004), social sorting (Lyon 2003), or improved interrogation techniques (McAllister 2017). Notwithstanding the above, companies can cause massive job losses due to AI implementation (Frey and Osborne 2013), conduct unmonitored forms of AI experiments on society without informed consent (Kramer et al. 2014), suffer from data breaches (Schneier 2018), use unfair, biased algorithms (Eubanks 2018), provide unsafe AI products (Sitawarin et al. 2018), use trade secrets to disguise harmful or flawed AI functionalities (Whittaker et al. 2018), rush to integrate and put immature AI applications on the market and many more. Furthermore, criminal or black-hat hackers can use AI to tailor cyberattacks, steal information, attack IT infrastructures, rig elections, spread misinformation for example through deepfakes, use voice synthesis technologies for fraud or social engineering (Bendel 2017), or disclose personal traits that are actually secret or private via machine learning applications (Kosinski and Wang 2018; Kosinski et al. 2013, 2015). All in all, only a very small number of papers is published about the misuse of AI systems, even though they impressively show what massive damage can be done with those systems (Brundage et al. 2018; King et al. 2019; O’Neil 2016).

While the United States currently has the largest number of start-ups, China claims to be the “world leader in AI” in 2030 (Abacus 2018). This claim is supported by the sheer amount of data that China has at its disposal to train its own AI systems, as well as by the large label companies that take over the manual preparation of data sets for supervised machine learning (Yuan 2018). Conversely, China is seen to have a weakness vis-à-vis the USA in that the investments of the market leaders Baidu, Alibaba and Tencent are too application-oriented comprising areas such as autonomous driving, finance or home appliances, while important basic research on algorithm development, chip production or sensor technology is neglected (Hao 2019). The constant comparison between China, the USA and Europe renders the fear of being inferior to each other an essential motive for efforts in the research and development of artificial intelligence.

Another justification for competitive thinking is provided by the military context. If the own “team”, framed in a nationalist way, does not keep pace, so the consideration, it will simply be overrun by the opposing “team” with superior AI military technology. In fact, potential risks emerge from the AI race narrative, as well as from an actual competitive race to develop AI systems for technological superiority (Cave and ÓhÉigeartaigh 2018). One risk of this rhetoric is that “impediments” in the form of ethical considerations will be eliminated completely from research, development and implementation. AI research is not framed as a cooperative global project, but as a fierce competition. This competition affects the actions of individuals and promotes a climate of recklessness, repression, and thinking in hierarchies, victory and defeat. The race for the best AI, whether a mere narrative or a harsh reality, reduces the likelihood of the establishment of technical precaution measures as well as of the development of benevolent AI systems, cooperation, and dialogue between research groups and companies. Thus, the AI race stands in stark contrast to the idea of developing an “AI4people” (Floridi et al. 2018). The same holds true for the idea of an “AI for Global Good”, as was proposed at the 2017’s ITU summit, or the large number of leading AI researchers who signed the open letter of the “Future of Life Institute”, embracing the norm that AI should be used for prosocial purposes.

Despite the downsides, in less public discourses and in concrete practice, an AI race has long since established itself. Along with that development, in- and outgroup-thinking has intensified. Competitors are seen more or less as enemies or at least as threats against which one has to defend oneself. Ethics, on the other hand, in its considerations and theories always stresses the danger of an artificial differentiation between in- and outgroups (Derrida 1997). Constructed outgroups are subject to devaluation, are perceived de-individualized and in the worst case can become victims of violence simply because of their status as “others” (Mullen and Hu 1989; Vaes et al. 2014). I argue that only by abandoning such thinking in- and outgroups may the AI race be reframed into a global cooperation for beneficial and safe AI.

Do ethical guidelines bring about a change in individual decision-making regardless of the larger social context? In a recent controlled study, researchers critically reviewed the idea that ethical guidelines serve as a basis for ethical decision-making for software engineers (McNamara et al. 2018). In brief, their main finding was that the effectiveness of guidelines or ethical codes is almost zero and that they do not change the behavior of professionals from the tech community. In the survey, 63 software engineering students and 105 professional software developers were scrutinized. They were presented with eleven software-related ethical decision scenarios, testing whether the influence of the ethics guideline of the Association for Computing Machinery (ACM) (Gotterbarn et al. 2018) in fact influences ethical decision-making in six vignettes, ranging from responsibility to report, user data collection, intellectual property, code quality, honesty to customer to time and personnel management. The results are disillusioning: “No statistically significant difference in the responses for any vignette were found across individuals who did and did not see the code of ethics, either for students or for professionals.” (McNamara et al. 2018, 4).

Irrespective of such considerations on the microsociological level, the relative ineffectiveness of ethics can also be explained at the macrosociological level. Countless companies are eager to monetize AI in a huge variety of applications. This strive for a profitable use of machine learning systems is not primarily framed by value- or principle-based ethics, but obviously by an economic logic. Engineers and developers are neither systematically educated about ethical issues, nor are they empowered, for example by organizational structures, to raise ethical concerns. In business contexts, speed is everything in many cases and skipping ethical considerations is equivalent to the path of least resistance. Thus, the practice of development, implementation and use of AI applications has very often little to do with the values and principles postulated by ethics. The German sociologist Ulrich Beck once stated that ethics nowadays “plays the role of a bicycle brake on an intercontinental airplane” (Beck 1988, 194). This metaphor proves to be particularly true in the context of AI, where huge sums of money are invested in the development and commercial utilization of systems based on machine learning (Rosenberg 2017), while ethical considerations are mainly used for public relations purposes (Boddington 2017, 56).

In their AI Now 2017 Report, Kate Crawford and her team state that ethics and forms of soft governance “face real challenges” (Campolo et al. 2017, 5). This is mainly due to the fact that ethics has no enforcement mechanisms reaching beyond a voluntary and non-binding cooperation between ethicists and individuals working in research and industry. So what happens is that AI research and development takes place in “closed-door industry settings”, where “user consent, privacy and transparency are often overlooked in favor of frictionless functionality that supports profit-driven business models” (Campolo et al. 2017, 31 f.). Despite this dispensation of ethical principles, AI systems are used in areas of high societal significance such as health, police, mobility or education. Thus, in the AI Now Report 2018, it is repeated that the AI industry “urgently needs new approaches to governance”, since, “internal governance structures at most technology companies are failing to ensure accountability for AI systems” (Whittaker et al. 2018, 4). Thus, ethics guidelines often fall into the category of a “’trust us’ form of [non-binding] corporate self-governance” (Whittaker et al. 2018, 30) and people should “be wary of relying on companies to implement ethical practices voluntarily” (Whittaker et al. 2018, 32).

The tension between ethical principles and wider societal interests on the one hand, and research, industry, and business objectives on the other can be explained with recourse to sociological theories. Especially on the basis of system theory it can be shown that modern societies differ in their social systems, each working with their own codes and communication media (Luhmann 1984, 1997, 1988). Structural couplings can lead decisions in one social system to influence other social systems. Such couplings, however, are limited and do not change the overall autonomy of social systems. This autonomy, which must be understood as an exclusive, functionalist orientation towards the system’s own codes is also manifested in the AI industry, business and science. All these systems have their own codes, their own target values, and their own types of economic or symbolic capital via which they are structured and based upon which decisions are made (Bourdieu 1984). Ethical intervention in those systems is only possible to a very limited extent (Hagendorff 2016). A certain hesitance exists towards every kind of intervention as long as these lie beyond the functional laws of the respective systems. Despite that, unethical behavior or unethical intentions are not solely caused by economic incentives. Rather, individual character traits like cognitive moral development, idealism, or job satisfaction play a role, let alone organizational environment characteristics like an egoistic work climate or (non-existent) mechanisms for the enforcement of ethical codes (Kish-Gephart et al. 2010). Nevertheless, many of these factors are heavily influenced by the overall economic system logic. Ethics is then, so to speak, “operationally effectless” (Luhmann 2008).

And yet, such system-theoretical considerations apply only on a macro level of observation and must not be generalized. Deviations from purely economic behavioral logics in the tech industry occur as well, for example when Google withdrew from the military project “Maven” after protests from employees (Statt 2018) or when people at Microsoft protested against the company’s cooperation with Immigration and Customs Enforcement (ICE) (Lecher 2018). Nevertheless, it must also be kept in mind here that, in addition to genuine ethical motives, the significance of economically relevant reputation losses should not be underestimated. Hence, the protest against unethical AI projects can in turn be interpreted in an economic logic, too.

As indicated in the previous sections, the practice of using AI systems is poor in terms of compliance with the principles set out in the various ethical guidelines. Great progress has been made in the areas of privacy, fairness or explainability. For example, many privacy-friendly techniques for the use of data sets and learning algorithms have been developed, using methods where AI systems’ “sight” is “darkened” via cryptography, differential or stochastic privacy (Ekstrand et al. 2018; Baron and Musolesi 2017; Duchi et al. 2013; Singla et al. 2014). Nevertheless, this contradicts the observation that AI has been making such massive progress for several years precisely because of the large amounts of (personal) data available. Those data are collected by privacy-invasive social media platforms, smartphone apps, as well as Internet of Things devices with its countless sensors. In the end, I would argue that the current AI boom coincides with the emergence of a post-privacy society. In many respects, however, this post-privacy society is also a black box society (Pasquale 2015), in which, despite technical and organizational efforts to improve explainability, transparency and accountability, massive zones of non-transparency remain, caused both by the sheer complexity of technological systems and by strategic organizational decisions.

For many of the issues mentioned in the guidelines, it is difficult to assess the extent to which efforts to meet the set objectives are successful or whether conflicting trends prevail. This is the case in the areas of safety and cybersecurity, the science-policy link, future of employment, public awareness about AI risks, or human oversight. In other areas, including the issue of hidden costs and sustainability, the protection of whistleblowers, diversity in the field of AI, the fostering of solidarity and social cohesion, the respect for human autonomy, the use of AI for the common good or the military AI arms race, it can certainly be stated that the ethical goals are being massively underachieved. One only has to think of the aspect of gender diversity: Even though ethical guidelines clearly demand its improvement, the state of affairs is that on average 80% of the professors at the world’s leading universities such as Stanford, Oxford, Berkeley or the ETH are male (Shoham et al. 2018). Furthermore, men make up more than 70% of applicants for AI jobs in the U.S. (Shoham et al. 2018). Alternatively, one can take human autonomy: As repeatedly demanded in various ethical guidelines, people should not be treated as mere data subjects, but as individuals. In fact, however, countless examples show that computer decisions, regardless of their susceptibility to error, are attributed a strong authority which results in the ignorance of individual circumstances and fates (Eubanks 2018). Furthermore, countless companies strive for the opposite of human autonomy, employing more and more subtle techniques for manipulating user behavior via micro targeting, nudging, UX-design and so on (Fogg 2003; Matz et al. 2017). Another example is that of cohesion: Many of the major scandals of the last years would have been unthinkable without the use of AI. From echo chamber effects (Pariser 2011) to the use of propaganda bots (Howard and Kollanyi 2016), or the spread of fake-news (Vosoughi et al. 2018), AI always played a key role to the effect of diminishing social cohesion, fostering instead radicalization, the decline of reason in public discourse and social divides (Tufekci 2018; Brady et al. 2017).

Given the relative lack of tangible impact of the normative objectives set out in the guidelines, the question arises as to how the guidelines could be improved to make them more effective. At first glance, the most obvious potential for improvement of the guidelines is probably to supplement them with more detailed technical explanations—if such explanations can be found. Ultimately, it is a major problem to deduce concrete technological implementations from the very abstract ethical values and principles. What does it mean to implement justice or transparency in AI-systems? What does a “human-centered” AI look like? How can human oversight be obtained? The list of questions could easily be continued.

The ethics guidelines examined refer exclusively to the term “AI”. They never or very seldom use more specific terminology. However, “AI” is just a collective term for a wide range of technologies or an abstract large-scale phenomenon. The fact that not a single prominent ethical guideline goes into greater technical detail shows how deep the gap is between concrete contexts of research, development, and application on the one side, and ethical thinking on the other. Ethicists must partly be capable of grasping technical details with their intellectual framework. That means reflecting on the ways data are generated, recorded, curated, processed, disseminated, shared, and used (Bruin and Floridi 2017), on the ways of designing algorithms and code, respectively (Kitchin 2017; Kitchin and Dodge 2011), or on the ways training data sets are selected (Gebru et al. 2018). In order to analyze all this in sufficient depth, ethics has to partially transform to “microethics”. This means that at certain points, a substantial change in the level of abstraction has to happen insofar as ethics aims to have a certain impact and influence in the technical disciplines and the practice of research and development of artificial intelligence (Morley et al. 2019). On the way from ethics to “microethics”, a transformation from ethics to technology ethics, to machine ethics, to computer ethics, to information ethics, to data ethics has to take place. As long as ethicists refrain from doing so, they will remain visible in a general public, but not in professional communities.

A good example of such a microethical work which can be implemented easily and concretely in practice is the paper by Gebru et al. (2018). The researchers propose the introduction of standardized datasheets listing the properties of different training data sets, so that machine learning-practitioners can check to what extent certain data sets are best suitable for their purposes, what the original intention was when the data set was created, what data the data set is composed of, how the data was collected and pre-processed, etc. The paper by Gebru et al. makes it possible for practitioners to obtain a more informed decision on the selection of certain training data sets, so that supervised machine learning ultimately becomes fairer, and more transparent, and avoids cases of algorithmic discrimination (Buolamwini and Gebru 2018). Such work is, however, an exception.

In general, ethical guidelines postulate very broad, overarching principles which are then supposed to be implemented in a widely diversified set of scientific, technical and economic practices, and in sometimes geographically dispersed groups of researchers and developers with different priorities, tasks and fragmental responsibilities. Ethics thus operates at a maximum distance from the practices it actually seeks to govern. Of course, this does not remain unnoticed among technology developers. In consequence, the generality and superficiality of ethical guidelines in many cases not only prevents actors from bringing their own practice into line with them, but rather encourages the devolution of ethical responsibility to others.

Regardless of the fact that normative guidelines should be accompanied by in-depth technical instructions—as far as they can reasonably be identified—, the question still arises how the precarious situation regarding the application and fulfillment of AI ethics guidelines can be improved. To address this question, one needs to take a step back and look at ethical theories in general. In ethics, several major strands of theories were created and shaped by various philosophical traditions. Those theories range from deontological to contractualistic, utilitarian, or virtue ethical approaches (Kant 1827; Rawls 1975; Bentham 1838; Hursthouse 2001). In the following, two of these approaches—deontology and virtue ethics—will be selected to illustrate different approaches in AI ethics. The deontological approach is based on strict rules, duties or imperatives. The virtue ethics approach, on the other hand, is based on character dispositions, moral intuitions or virtues—especially “technomoral virtues” (Vallor 2016). In the light of these two approaches, the traditional type of AI ethics can be assigned to the deontological concept (Mittelstadt 2019). Ethics guidelines postulate a fixed set of universal principles and maxims which technology developers should adhere to (Ananny 2016). The virtue ethics approach, on the other hand, focuses more on “deeper-lying” structures and situation-specific deliberations, on addressing personality traits and behavioral dispositions on the part of technology developers (Leonelli 2016). Virtue ethics does not define codes of conduct but focusses on the individual level. The technologists or software engineers and their social context are the primary addressees of such an ethics (Ananny 2016), not technology itself.

I argue that the prevalent approach of deontological AI ethics should be augmented with an approach oriented towards virtue ethics aiming at values and character dispositions. Ethics is then no longer understood as a deontologically inspired tick-box exercise, but as a project of advancing personalities, changing attitudes, strengthen responsibilities and gaining courage to refrain from certain actions, which are deemed unethical. When following the path of virtue ethics, ethics as a scientific discipline must refrain from wanting to limit, control, or steer (Luke 1995). Very often, ethics or ethical guidelines are perceived as something whose purpose is to stop or prohibit activity, to hamper valuable research and economic endeavors (Boddington 2017, 8). I want to resign this negative notion of ethics. It should not be the objective of ethics to stifle activity, but to do the exact opposite, i.e. broadening the scope of action, uncovering blind spots, promoting autonomy and freedom, and fostering self-responsibility.

In view of AI ethics, approaches that focus on virtues aim at cultivating a moral character, expressing technomoral virtues such as honesty, justice, courage, empathy, care, civility, or magnanimity, to name just a few (Vallor 2016). Those virtues are supposed to raise the likelihood of ethical decision-making practices in organizations that develop and deploy AI applications. Cultivating a moral character, in terms of virtue ethics, means to educate virtues in families, schools, communities, as well as companies. At best, every individual, every member of a society should encourage this cultivation, by generating the motivation to adopt and habituate practices that influence technology development and use in a positive manner. Especially the subject of responsibility diffusion can only be circumvented when virtue ethics is adopted on a broad and collective level in communities of tech professionals. Simply every person involved in data science, data engineering and data economies related to applications of AI has to take at least some responsibility for the implications of their actions (Leonelli 2016). This is why researchers such as Floridi argue that every actor who is causally relevant for bringing about the collective consequence or impacts in question, has to be held accountable (Floridi 2016). Interestingly, Floridi uses the backpropagation method known from Deep Learning to describe the way in which responsibilities can be assigned, except that here backpropagation is used in networks of distributed responsibility. When working in groups, actions that are on first glance allegedly morally neutral can nevertheless have consequences or impacts—intended or non-intended—that are morally wrong. This means that practitioners from AI communities always need to discern the overarching, short- and long-term consequences of the technical artefacts they are building or maintaining, as well as to explore alternative ways of developing software or using data, including the option of completely refraining from carrying out particular tasks, which are considered unethical.

In addition to the endorsement of virtue ethics in tech communities, several institutional changes should take place. They include the adoption of legal framework conditions, the establishment of mechanisms for an independent auditing of technologies, the establishment of institutions for complaints, which also compensate for harms caused by AI systems, and the expansion of university curricula in particular through content from ethics of technology, media, and information (Floridi et al. 2018; Cowls and Floridi 2018; Eaton et al. 2017; Goldsmith and Burton 2017). So far, however, hardly any of these demands have been met.