AI Ethics Throughout the Years*

 * This text is retrieved from my dissertation Empirical Essays on Artificial Intelligence Ethics

Normative discussions about AI have run parallel to its development throughout the years. These discussions reflect the achievements, failures, and societal anxieties associated with these technologies. Some of the issues explored in the past, related to explainability, accountability, or privacy, remain relevant today.

The Early Days

In the early days of AI, the focus was on getting computers to do things that would be regarded as intelligent if done by humans [1]. The philosophical discussions that took place around this time were not so much about morality but rather about epistemological questions concerning minds and machines: Can machines think?, Can machines be intelligent?, Can an artificial brain potentially outperform a human brain? [1, 2]. These foundational questions would continue to be explored throughout the following decades, but the philosophical debate became more diverse as AI evolved.

Some moral considerations surfaced in the literature of the 1960s [3]. At this time, reckless claims were made regarding the potential of AI to copy and perhaps even recreate the entire workings of the human brain within a short timeline [1]. Those claims prompted concerns about AI technology. The Wiener-Samuel debate illustrates early discussions about the potential for machines to threaten humankind. Norbert Wiener famously wrote that machines can and do transcend some of the limitations of their designers, and in doing so they may be both effective and dangerous [4]. Arthur Samuel disagreed on the grounds that machines could not do anything unless instructed by humans [5]. He conceded, however, that projected machines of the so-called "neural net" type could be an exception to the deterministic nature of machines. Since their internal connections would be unknown, the precise behavior of the nets would be unpredictable and, therefore, potentially dangerous [5]. Ahead of his time, Samuel was alluding to what is known today as the problem of explainability associated with sub-symbolic AI algorithms. 

The ethical and societal implications of AI were initially explored in the 1970s. This was a period of great disenchantment with AI, as the technology had failed to meet expectations and deliver the promised results [1]. In addition to the technical frustrations and lack of funding, important philosophical work challenging the claims of AI was published around this time. Hubert Dreyfus published What Computers Can’t do, in which he argued that machines could not display higher mental functions through the use of symbolic representations [6, 7]. John Searle came up with his Chinese room argument to show that a computer could not be said to understand the symbols with which it communicates [8]. Eventually, the claims about AI were toned down. Many researchers followed the lead of John McCarthy to develop AI programs limited to a particular domain of knowledge with practical applications in industry [1, 9].

Around this period, a Delphi study was conducted among experts in the AI field with the purpose of understanding both the capabilities and limitations of machine intelligence and its potential impact on society [10]. This study accurately predicted the use of AI in mundane activities such as domestic chores, entertainment, or weather forecast and addressed ethical issues related to safety, privacy, and jobs displacement. It was also cautioned that there was a need to formalize algorithmically some of the ethical and empirical rules and trade-offs that society observes implicitly [10]: It may be necessary for a robot to have available rules which dictate how to trade-off life for property, e.g., when is it allowable to wreck an automobile to avoid killing an animal? [10]. This sort of moral dilemma would eventually become the focal point of AI Ethics decades later.

Accountability of Expert Systems, Singularity & Asimov

Following the lead of McCarthy, throughout the 1980s and 1990s, the AI community focused on the development of expert systems. These computer systems were designed to solve complex reasoning problems in particular domains at the level of performance of human experts [11]. The normative discussions during the heyday of expert systems were residual, although some scholars explored issues of moral responsibility and accountability associated with these systems [12–14]. Expert systems were developed to operate in different domains, such as Medicine and Law [15]. However, they facedmany challenges, mainly because the medical or legal daily practice can not be reduced to a closed set of rules [16, 17]. 

Despite only moderate success of Expert Systems, there was an ongoing conversation about technological singularity, i.e., a hypothetical context where super-intelligent machines design and produce even more super-intelligent machines [1, 18–20]. The Paperclip Maximizer introduced by Nick Bostrom, which describes the existential risk that superintelligence may pose to human beings when programmed to pursue seemingly harmless goals [20], is yet another illustration of the concerns over the potential of machines to threaten humankind. 

Contrasting with those early discussions in the 1960s, this time, the scientific community was interested in finding solutions for the problem of complex systems lacking morality [21]. In the absence of a moral theory, the three laws of robotics, introduced by science fiction writer Isaac Asimov, became a fixture in scientific publications about the ethics of machines and AI [1].

Machine Ethics, Trolley Problems, & Empirical Ethics

Concerns about the moral behavior of increasingly intelligent and autonomous machines led to the emergence ofMachine Ethics. This research field aims at equipping machines with ethical reasoning to ensure that their behavior towards humans and other machines is ethically acceptable [22–26]. The core idea is that sensitivity to ethics should be integral to the software of machines in order to facilitate their ethical use [24].

Unlike traditional philosophy of technology, which was primarily reactive, and modern philosophy of technology, which is proactive in raising awareness of the values designers bring to the technology design process,Machine Ethics went one step further, seeking to build ethical decision-making capacities directly into the machines [24, 27].

The need for artificial morality was reinforced with the advent of autonomous driving. The novel AMA was an AV in an extreme traffic situation. Upon a string of high profile publications [28, 29], the AV trolley problem, which had been timidly addressed in the 1970s Delphi study, was heavily featured in the scientific and popular literature [72, 98, 30–34]. The discussions revolved around the practical relevance of the AV moral dilemma [35–37], the merits of using different ethical frameworks as control algorithms for AVs [38–43], and the moral preferences and societal expectations about the ethics to be encoded in AVs [28, 29, 44, 45]. It is not unlikely that one day the scientific community will disapprove of the amount of attention paid to the AV trolley. However, it is indisputable that such attention played a crucial role in raising awareness about AI Ethics.

Guidelines & Principles

Increased awareness about AI Ethics led to a proliferation of soft governance mechanisms for the ethical development and deployment of AI [46, 47]. Many organizations have launched a wide range of initiatives, such as codes, guidelines, frameworks, and policy strategies, to establish ethical principles for the adoption of socially beneficial AI [48]. These initiatives brought the focus of the AI Ethics debates to a common set of issues and principles. Although varying in terminology, it is reported that the different guidelines broadly converge around five principles: (i) beneficence, (ii) non-maleficence, (iii) autonomy, (iv) justice, and (v) explicability [46, 48, 49]. 

Ethically-aligned AI should therefore be beneficial to people and the environment (beneficence); robust and secure (non-maleficence); respectful of human values (autonomy); fair (justice); and explainable and accountable (explicability). These principles closely resemble the four classic principles in medical ethics [50]. The association is convenient since medical ethics is historically the most prominent and well-studied approach to applied ethics, however it may not be warranted due to important differences between the two fields [51] . 

However, because these principles remain abstract and are not translated to practice, the AI Ethics endeavor may be falling short of meeting its goals of providing normative guidance for the design and deployment of algorithms [47, 53]. A poor translation of these principles and guidelines into practices leaves room for unethical behaviors such as ethics shopping, i.e., mixing and matching ethical principles from different sources to justify some pre-existing behavior [54]; ethics bluewashing, i.e., making unsubstantiated or misleading claims about the efforts and resources allocated to address ethical problems associated with AI; ethics dumping, i.e., exporting research activities about digital processes, products, services, or other solutions, in other contexts or places in ways that would be ethically unacceptable in the context or place of origin and importing the outcomes of such unethical research activities[54]; ethics shirking, i.e., doing increasingly less ethical work in a given context, thus lowering the overall level of ethics engagement [54]; and ethics lobbying, i.e., exploiting ethics to delay or avoid good and necessary legislation [54]. Concerns about these unethical behaviors related to AI are markedly present in the recent and more practical work developed in AI Ethics.

AI Ethics today

To a great extent, the normative work developed in previous decades has shaped the field of AI Ethics. Research carried out in AI Ethics ranges from (i) reflections and practical work on how ethical principles can be implemented in decision routines of autonomous machines (Machine Ethics) [41, 55–57]; (ii) empirical analysis on how moral dilemmas are solved (Empirical AI Ethics) [28, 29]; and (iii) comprehensive AI principles and guidelines (Principles & Guidelines)[46]. 

Despite the richness of the AI Ethics work, a growing body of the literature has declared that it is failing to realize its normative endeavors. There is little evidence on the plausibility of the Machine Ethics project, as AMAs remain, for now, proofs of concept and lab prototypes [57]. The empirical AI Ethics studies have provided rich information about the societal expectations and preferences in moral situations involving AI. However, some of these studies, such as theMoralMachine Experiment (MME), have been criticized for their stylized and unrealistic premises (moral dilemmas) [29, 58]. The impact of the AI Ethics guidelines developed to promote ethical practices in AI is also disputed. There is a mistrust that organizations, particularly those involved in the development of AI technologies, will implement ethical practices voluntarily [59]. When Ethics is integrated into organizations, there are concerns that it is used merely as a marketing strategy with little impact when it comes to decisions made in the AI domain [47, 60, 61]. 

A study that surveyed AI practitioners about their perceived impact of AI Ethics guidelines reported that the effectiveness of such guidelines or ethical codes is almost zero and that they do not change the behavior of professionals fromthe tech community [60]. The translational work aiming at producing tools or methods for implementing ethics into practice has also been scrutinized. There is little evidence of the impact of existing translational tools on the governance of AI [91]. Morley et al. identified the tools and methods already available to guide AI practitioners on core issues of AI Ethics and plotted these methods and tools in a typology, matching them to ethical principles (beneficence, non-maleficence, autonomy, justice, and explicability) to stages in the algorithm development pipeline. They reported that numerous tools and methodologies exist to assist practitioners in realizing Ethical AI, but the vast majority are severely limited in terms of usability [91]. 

Poor guidance on how to produce Ethical AI may result in significant societal opportunity costs [91]. AI is a multi-purpose technology with the potential to improve human well being. However, consumers may be discouraged from adopting these technologies if the costs of ethical mistakes outweigh the benefits of ethical success. Weak public acceptance means slimmer chances for AI to meet its potential. Resorting once again to the archetypal AV case, it is clear that the ethical trade-offs in traffic situations that entail distribution of risk will need to be resolved for this technology to be accepted by the public, thusmeeting its potential to save lives [28, 58, 62]. 

The so-called failure of AI Ethics to realize its normative mandate raises unsettling concerns about the feasibility of this endeavor. Some of those concerns may be unwarranted as stemming from a misconception about Ethics. It is not expected that AI Ethics provides policies or regulations but rather a operational normative framework. Past experience with other applied fields of Ethics (e.g., medical ethics) shows that it is possible to operationalize Ethics successfully, or at least abstract ethical principles, thus lending credence to the AI Ethics effort [53]. It is hypothesized that the current struggles of AI Ethics may be rooted in the challenges that have emerged in the current data-driven paradigm in AI and that empirical work may be valuable in overcoming those challenges.

References:

[1] K.Warwick, Artificial intelligence: the basics, edited by Routledge (Routledge, 2013).

[2] A. M. Turing and J. Haugeland, Computing machinery and intelligence (MIT Press Cambridge, MA, 1950).

[3] N. Wiener, God and Golem, Inc: A Comment on Certain Points where Cybernetics Impinges on Religion, Vol. 42 (MIT press, 1964).

[4] N.Wiener, Some moral and technical consequences of automation, Science 26, 305 (1960).

[5] A. L. Samuel, Some moral and technical consequences of automation—a refutation, Science 27, 741 (1960).

[6] H. L. Dreyfus, What computers can’t do : a critique of artificial reason (Harper & Row, 1972).

[7] H. L. Dreyfus, L. Hubert, et al., What computers still can’t do: A critique of artificial reason (MIT press, 1992).

[8] J. R. Searle,Minds, brains, and programs, Behavioral and Brain Sciences 3, 417–424 (1980).

[9] J.McCarthy, Generality in artificial intelligence, Commun. ACM 30, 1030–1035 (1987).

[10] O. Firschein, M. A. Fischler, L. S. Coles, and J. M. Tenenbaum, Forecasting and assessing the impact of artificial intelligence on society, in IJCAI, Vol. 5 (Citeseer, 1973) pp. 105–15.

[11] E. A. Feigenbaum, Knowledge Engineering: The Applied Side of Artificial Intelligence., Tech. Rep. (STANFORD UNIV CA DEPT OF COMPUTER SCIENCE, 1980).

[12] G. Brown and K. Kornmayer, Expert systems restructure managed care practice: implementation and ethics, Behavioral healthcare tomorrow 5, 31—34 (1996).

[13] K. Cass, Expert systems as general-use advisory tools: An examination of moral responsibility, Business and Professional Ethics Journal 15, 61 (1996).

[14] J. F. Dillard and K. Yuthas, A responsibility-based approach to the development of expert systems, in Proceedings of the International Association for Business and Society, Vol. 7 (1996) pp. 351–360.

[15] Shu-Hsien Liao, Expert system methodologies and applications—a decade review from 1995 to 2004, Expert Systems with Applications 28, 93 (2005).

[16] H. Heathfield, The rise and ‘fall’ of expert systems in medicine, Expert Systems 16, 183 (1999), https://onlinelibrary.wiley.com/doi/pdf/10.61/418-0394.002 .

[17] P. Leith, The rise and fall of the legal expert system, International Review of Law, Computers & Technology 30, 94 (2016), https://doi.org/10.30/3100869.2016.182465

[18] H. Moravec, Mind children: The future of robot and human intelligence (Harvard University Press, 1988).

[19] V. Vinge, Technological singularity, in VISION-21 Symposium sponsored by NASA Lewis Research Center and the Ohio Aerospace Institute (1993) pp. 30–31.

[20] N. Bostrom, Ethical issues in advanced artificial intelligence, Science fiction and philosophy: from time travel to superintelligence , 277 (2003).

[21] E. Yudkowsky, Creating friendly ai 1.0: The analysis and design of benevolent goal architectures, The Singularity Institute, San Francisco, USA (2001).

[22] M. Anderson, S. L. Anderson, and C. Armen, Towards machine ethics, in AAAI-04 workshop on agent organizations: theory and practice, San Jose, CA (2004).

[23] J. H.Moor, The nature, importance, and difficulty of machine ethics, IEEE intelligent systems 21, 18 (2006).

[24] C. Allen, I. Smit, and W. Wallach, Artificial morality: Top-down, bottom-up, and hybrid approaches, Ethics and information technology 7, 44 (2005).

[25] M. Anderson and S. L. Anderson, The status of machine ethics: a report from the aaai symposium,Minds andMachines 17, 1 (2007).

[26] M. Anderson and S. L. Anderson,Machine ethics: Creating an ethical intelligent agent, inMachine Ethics and Robot Ethics (Routledge, 2020) pp. 237–248.

[27] C. Allen,W.Wallach, and I. Smit, Why machine ethics? IEEE Intelligent Systems 21, 12 (2006).

[28] J.-F. Bonnefon, A. Shariff, and I. Rahwan, The social dilemma of autonomous vehicles, Science 352, 523 (2016).

[29] E. Awad, S. Dsouza, R. Kim, J. Schulz, J. Henrich, A. Shariff, J.-F. Bonnefon, and I. Rahwan, The moral machine experiment, Nature 563, 59 (2018).

[30] N. J. Goodall,Machine ethics and automated vehicles, in Road Vehicle Automation, edited by G.Meyer and S. Beiker (Springer International Publishing, Cham, 2014) pp.93–102.

[31] L. Greenemeier, Driverless cars will face moral dilemmas, Scientific American (2016).

[32] F. S. de Sio, Killing by autonomous vehicles and the legal doctrine of necessity, Ethical Theory andMoral Practice 20, 411 (2017).

[33] R. Sparrow and M. Howard, When human beings are like drunk robots: Driverless vehicles, ethics, and the future of transport, Transportation Research Part C: Emerging Technologies 80, 206 (2017).

[34] N. JafariNaimi, Our bodies in the trolley’s path, or why self-driving cars must *not* be programmed to kill, Science, Technology, & Human Values 43, 302 (2018), https://doi.org/10.22/057243917718942 .

[35] J.Himmelreich,Never mind the trolley: The ethics of autonomous vehicles inmundane situations, Ethical Theory andMoral Practice , 1 (2018).

[36] G. Keeling, Why trolley problems matter for the ethics of automated vehicles, Science and engineering ethics , 1 (2019).

[37] A.Wolkenstein, What has the trolley dilemma ever done for us (and what will it do in the future)? on some recent debates about the ethics of self-driving cars, Ethics and Information Technology , 1 (2018).

[38] J. C. Gerdes and S. M. Thornton, Implementable ethics for autonomous vehicles, in Autonomes Fahren: Technische, rechtliche und gesellschaftliche Aspekte, edited by M.Maurer, J. C. Gerdes, B. Lenz, and H.Winner (Springer Berlin Heidelberg, Berlin, Heidelberg, 2015) pp. 87–102.

[39] P. Lin, Why ethics matters for autonomous cars, in Autonomous driving (Springer, Berlin, Heidelberg, 2016) pp. 69–85.

[40] D. Leben, A rawlsian algorithm for autonomous vehicles, Ethics and Information Technology 19, 2 (2017).

[41] S.M. Thornton, S. Pan, S. M. Erlien, and J. C. Gerdes, Incorporating ethical considerations into automated vehicle control, IEEE Transactions on Intelligent Transportation Systems 18, 379 (2017).

[42] L. T. Bergmann, L. Schlicht, C. Meixner, P. König, G. Pipa, S. Boshammer, and A. Stephan, Autonomous vehicles require socio-political acceptance—an empirical and philosophical perspective on the problem of moral decision making, Frontiers in behavioral neuroscience 12, 31 (2018).

[43] M. Cunneen,M.Mullins, F.Murphy, D. Shannon, I. Furxhi, and C. Ryan, Autonomous vehicles and avoiding the trolley (dilemma): Vehicle perception, classification, and the challenges of framing decision ethics, Cybernetics and Systems 51, 59 (2020), https://doi.org/10.30/01969722.2019.610541 .

[44] E. Awad, S. Dsouza, A. Shariff, I. Rahwan, and J.-F. Bonnefon, Universals and variations in moral decisions made in 42 countries by 70,000 participants, Proceedings of the National Academy of Sciences 12, 2332 (2020).

[45] Y. E. Bigman and K. Gray, Life and death decisions of autonomous vehicles, Nature 579, E1 (2020).

[46] A. Jobin, M. Ienca, and E. Vayena, The global landscape of ai ethics guidelines, Nature Machine Intelligence 1, 389 (2019).

[47] T.Hagendorff, The ethics of ai ethics: An evaluation of guidelines,Minds andMachines , 1 (2020).

[48] L. Floridi and J. Cowls, A unified framework of five principles for ai in society, Available at SSRN 383271 (2019).

[49] J. Mökander, J.Morley,M. Taddeo, and L. Floridi, Ethics-based auditing of automated decision-making systems: Nature, scope, and limitations, Science and Engineering Ethics 27, 44 (2021).

[50] A. Jobin, M. Ienca, and E. Vayena, The global landscape of ai ethics guidelines, Nature Machine Intelligence 1, 389 (2019).

[51] B.Mittelstadt, Principles alone cannot guarantee ethical ai, NatureMachine Intelligence 1, 501 (2019).

[52] D. Schiff, J. Biddle, J. Borenstein, and K. Laas, What’s next for ai ethics, policy, and governance? a global overview, in Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, AIES ’20 (Association for ComputingMachinery, New York, NY, USA, 2020) p. 48–53.

[53] J.Morley, A. Elhalal, F. Garcia, L. Kinsey, J.Mökander, and L. Floridi, Ethics as a service: A pragmatic operationalisation of ai ethics,Minds andMachines 31, 239 (2021).

[54] L. Floridi, Translating principles into practices of digital ethics: Five risks of being unethical, Philosophy & Technology 32, 185 (2019).

[55] M. Anderson and S. Anderson,Machine Ethics (Cambridge University Press, 2011).

[56] A. Etzioni and O. Etzioni, Incorporating ethics into artificial intelligence, The Journal of Ethics 21, 403 (2017).

[57] J.-A. Cervantes, S. López, L.-F. Rodríguez, S. Cervantes, F. Cervantes, and F. Ramos, Artificial moral agents: A survey of the current status, Science and Engineering Ethics 26, 501 (2020).

[58] J.-F. Bonnefon, A. Shariff, and I. Rahwan, The trolley, the bull bar, and why engineers should care about the ethics of autonomous cars [point of view], Proceedings of the IEEE 2, 502 (2019).

[59] M. Whittaker, K. Crawford, R. Dobbe, G. Fried, E. Kaziunas, V. Mathur, S. M. West, R. Richardson, J. Schultz, and O. Schwartz, AI now report 2018 (AI Now Institute at New York University New York, 2018).

[60] A.McNamara, J. Smith, and E.Murphy-Hill, Does acm’s code of ethics change ethical decision making in software development? in Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, ESEC/FSE 2018 (Association for Computing Machinery, New York, NY, USA, 2018) p. 729–733.

[61] W. Orr and J. L. Davis, Attributions of ethical responsibility by artificial intelligence practitioners, Information, Communication & Society 23, 719 (2020),

https://doi.org/10.30/31913X.2020.173342

[62] A. Martinho, N. Herber, M. Kroesen, and C. Chorus, Ethical issues in focus by the autonomous vehicles industry, Transport reviews 41, 556 (2021).