AI4People is not the first initiative to consider the ethical implications of AI. Many organisations have already produced statements of the values or principles that should guide the development and deployment of AI in society. Rather than conduct a similar, potentially redundant exercise here, we strive to move the dialogue forward, constructively, from principles to proposed policies, best practices, and concrete recommendations for new strategies. Such recommendations are not offered in a vacuum. But rather than generating yet another series of principles to serve as an ethical foundation for our recommendations, we offer a synthesis of existing sets of principles produced by various reputable, multi-stakeholder organisations and initiatives. A fuller explanation of the scope, selection and method of assessing these sets of principles is available in Cowls and Floridi (Forthcoming). Here, we focus on the commonalities and noteworthy differences observable across these sets of principles, in view of the 20 recommendations offered in the rest of the paper. The documents we assessed are:
The Asilomar AI Principles, developed under the auspices of the Future of Life Institute, in collaboration with attendees of the high-level Asilomar conference of January 2017 (hereafter “Asilomar”; Asilomar AI Principles 2017);
The Montreal Declaration for Responsible AI, developed under the auspices of the University of Montreal, following the Forum on the Socially Responsible Development of AI of November 2017 (hereafter “Montreal”; Montreal Declaration 2017)Footnote 3;
The General Principles offered in the second version of Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems. This crowd-sourced global treatise received contributions from 250 global thought leaders to develop principles and recommendations for the ethical development and design of autonomous and intelligent systems, and was published in December 2017 (hereafter “IEEE”; IEEE 2017)Footnote 4;
The Ethical Principles offered in the Statement on Artificial Intelligence, Robotics and ‘Autonomous’ Systems, published by the European Commission’s European Group on Ethics in Science and New Technologies, in March 2018 (hereafter “EGE”; EGE 2018);
The “five overarching principles for an AI code” offered in paragraph 417 of the UK House of Lords Artificial Intelligence Committee’s report, AI in the UK: ready, willing and able?, published in April 2018 (hereafter “AIUK”; House of Lords 2018); and
The Tenets of the Partnership on AI, a multistakeholder organisation consisting of academics, researchers, civil society organisations, companies building and utilising AI technology, and other groups (hereafter “the Partnership”; Partnership on AI 2018).
Taken together, they yield 47 principles.Footnote 5 Overall, we find an impressive and reassuring degree of coherence and overlap between the six sets of principles. This can most clearly be shown by comparing the sets of principles with the set of four core principles commonly used in bioethics: beneficence, non-maleficence, autonomy, and justice. The comparison should not be surprising. Of all areas of applied ethics, bioethics is the one that most closely resembles digital ethics in dealing ecologically with new forms of agents, patients, and environments (Floridi 2013). The four bioethical principles adapt surprisingly well to the fresh ethical challenges posed by artificial intelligence. But they are not exhaustive. On the basis of the following comparative analysis, we argue that one more, new principle is needed in addition: explicability, understood as incorporating both intelligibility and accountability.
Beneficence: Promoting Well-Being, Preserving Dignity, and Sustaining the Planet
Of the four core bioethics principles, beneficence is perhaps the easiest to observe across the six sets of principles we synthesise here. The principle of creating AI technology that is beneficial to humanity is expressed in different ways, but it typically features at the top of each list of principles. Montreal and IEEE principles both use the term “well-being”: for Montreal, “the development of AI should ultimately promote the well-being of all sentient creatures”; while IEEE states the need to “prioritize human well-being as an outcome in all system designs”. AIUK and Asilomar both characterise this principle as the “common good”: AI should “be developed for the common good and the benefit of humanity”, according to AIUK. The Partnership describes the intention to “ensure that AI technologies benefit and empower as many people as possible”; while the EGE emphasises the principle of both “human dignity” and “sustainability”. Its principle of “sustainability” represents perhaps the widest of all interpretations of beneficence, arguing that “AI technology must be in line with … ensur[ing] the basic preconditions for life on our planet, continued prospering for mankind and the preservation of a good environment for future generations”. Taken together, the prominence of these principles of beneficence firmly underlines the central importance of promoting the well-being of people and the planet.
Non-maleficence: Privacy, Security and “Capability Caution”
Though “do only good” (beneficence) and “do no harm” (non-maleficence) seem logically equivalent, in both the context of bioethics and of the ethics of AI they represent distinct principles, each requiring explication. While they encourage well-being, the sharing of benefits and the advancement of the public good, each of the six sets of principles also cautions against the many potentially negative consequences of overusing or misusing AI technologies. Of particular concern is the prevention of infringements on personal privacy, which is listed as a principle in five of the six sets, and as part of the “human rights” principles in the IEEE document. In each case, privacy is characterised as being intimately linked to individuals’ access to, and control over, how personal data is used.
Yet the infringement of privacy is not the only danger to be avoided in the adoption of AI. Several of the documents also emphasise the importance of avoiding the misuse of AI technologies in other ways. The Asilomar Principles are quite specific on this point, citing the threats of an AI arms race and of the recursive self-improvement of AI, as well as the need for “caution” around “upper limits on future AI capabilities”. The Partnership similarly asserts the importance of AI operating “within secure constraints”. The IEEE document meanwhile cites the need to “avoid misuse”, while the Montreal Declaration argues that those developing AI “should assume their responsibility by working against the risks arising from their technological innovations”, echoed by the EGE’s similar need for responsibility.
From these various warnings, it is not entirely clear whether it is the people developing AI, or the technology itself, which should be encouraged not to do harm—in other words, whether it is Frankenstein or his monster against whose maleficence we should be guarding. Confused also is the question of intent: promoting non-maleficence can be seen to incorporate the prevention of both accidental (what we above call “overuse”) and deliberate (what we call “misuse”) harms arising. In terms of the principle of non-maleficence, this need not be an either/or question: the point is simply to prevent harms arising, whether from the intent of humans or the unpredicted behaviour of machines (including the unintentional nudging of human behaviour in undesirable ways). Yet these underlying questions of agency, intent and control become knottier when we consider the next principle.
Autonomy: The Power to Decide (Whether to Decide)
Another classic tenet of bioethics is the principle of autonomy: the idea that individuals have a right to make decisions for themselves about the treatment they do or not receive. In a medical context, this principle of autonomy is most often impaired when patients lack the mental capacity to make decisions in their own best interests; autonomy is thus surrendered involuntarily. With AI, the situation becomes rather more complex: when we adopt AI and its smart agency, we willingly cede some of our decision-making power to machines. Thus, affirming the principle of autonomy in the context of AI means striking a balance between the decision-making power we retain for ourselves and that which we delegate to artificial agents.
The principle of autonomy is explicitly stated in four of the six documents. The Montreal Declaration articulates the need for a balance between human- and machine-led decision-making, stating that “the development of AI should promote the autonomy of all human beings and control … the autonomy of computer systems” (italics added). The EGE argues that autonomous systems “must not impair [the] freedom of human beings to set their own standards and norms and be able to live according to them”, while AIUK adopts the narrower stance that “the autonomous power to hurt, destroy or deceive human beings should never be vested in AI”. The Asilomar document similarly supports the principle of autonomy, insofar as “humans should choose how and whether to delegate decisions to AI systems, to accomplish human-chosen objectives”.
These documents express a similar sentiment in slightly different ways, echoing the distinction drawn above between beneficence and non-maleficence: not only should the autonomy of humans be promoted, but also the autonomy of machines should be restricted and made intrinsically reversible, should human autonomy need to be re-established (consider the case of a pilot able to turn off the automatic pilot and regain full control of the airplane). Taken together, the central point is to protect the intrinsic value of human choice—at least for significant decisions—and, as a corollary, to contain the risk of delegating too much to machines. Therefore, what seems most important here is what we might call “meta-autonomy”, or a “decide-to-delegate” model: humans should always retain the power to decide which decisions to take, exercising the freedom to choose where necessary, and ceding it in cases where overriding reasons, such as efficacy, may outweigh the loss of control over decision-making. As anticipated, any delegation should remain overridable in principle (deciding to decide again).
The decision to make or delegate decisions does not take place in a vacuum. Nor is this capacity to decide (to decide, and to decide again) distributed equally across society. The consequences of this potential disparity in autonomy are addressed in the final of the four principles inspired by bioethics.
Justice: Promoting Prosperity and Preserving Solidarity
The last of the four classic bioethics principles is justice, which is typically invoked in relation to the distribution of resources, such as new and experimental treatment options or simply the general availability of conventional healthcare. Again, this bioethics principle finds clear echoes across the principles for AI that we analyse. The importance of “justice” is explicitly cited in the Montreal Declaration, which argues that “the development of AI should promote justice and seek to eliminate all types of discrimination”, while the Asilomar Principles include the need for both “shared benefit” and “shared prosperity” from AI. Under its principle named “Justice, equity and solidarity”, the EGE argues that AI should “contribute to global justice and equal access to the benefits” of AI technologies. It also warns against the risk of bias in datasets used to train AI systems, and—unique among the documents—argues for the need to defend against threats to “solidarity”, including “systems of mutual assistance such as in social insurance and healthcare”. The emphasis on the protection of social support systems may reflect geopolitics, insofar as the EGE is a European body. The AIUK report argues that citizens should be able to “flourish mentally, emotionally and economically alongside artificial intelligence”. The Partnership, meanwhile, adopts a more cautious framing, pledging to “respect the interests of all parties that may be impacted by AI advances”.
As with the other principles already discussed, these interpretations of what justice means as an ethical principle in the context of AI are broadly similar, yet contain subtle distinctions. Across the documents, justice variously relates to
Using AI to correct past wrongs such as eliminating unfair discrimination;
Ensuring that the use of AI creates benefits that are shared (or at least shareable); and
Preventing the creation of new harms, such as the undermining of existing social structures.
Notable also are the different ways in which the position of AI, vis-à-vis people, is characterised in relation to justice. In Asilomar and EGE respectively, it is AI technologies themselves that “should benefit and empower as many people as possible” and “contribute to global justice”, whereas in Montreal, it is “the development of AI” that “should promote justice” (italics added). In AIUK, meanwhile, people should flourish merely “alongside” AI. Our purpose here is not to split semantic hairs. The diverse ways in which the relationship between people and AI is described in these documents hints at broader confusion over AI as a man-made reservoir of “smart agency”. Put simply, and to resume our bioethics analogy, are we (humans) the patient, receiving the “treatment” of AI, the doctor prescribing it? Or both? It seems that we must resolve this question before seeking to answer the next question of whether the treatment will even work. This is the core justification for our identification within these documents of a new principle, one that is not drawn from bioethics.
Explicability: Enabling the Other Principles Through Intelligibility and Accountability
The short answer to the question of whether “we” are the patient or the doctor is that actually we could be either—depending on the circumstances and on who “we” are in our everyday life. The situation is inherently unequal: a small fraction of humanity is currently engaged in the design and development of a set of technologies that are already transforming the everyday lives of just about everyone else. This stark reality is not lost on the authors whose documents we analyse. In all, reference is made to the need to understand and hold to account the decision-making processes of AI. This principle is expressed using different terms: “transparency” in Asilomar; “accountability” in EGE; both “transparency” and “accountability” in IEEE; “intelligibility” in AIUK; and as “understandable and interpretable” for the Partnership. Though described in different ways, each of these principles captures something seemingly novel about AI: that its workings are often invisible or unintelligible to all but (at best) the most expert observers.
The addition of this principle, which we synthesise as “explicability” both in the epistemological sense of “intelligibility” (as an answer to the question “how does it work?”) and in the ethical sense of “accountability” (as an answer to the question: “who is responsible for the way it works?”), is therefore the crucial missing piece of the jigsaw when we seek to apply the framework of bioethics to the ethics of AI. It complements the other four principles: for AI to be beneficent and non-maleficent, we must be able to understand the good or harm it is actually doing to society, and in which ways; for AI to promote and not constrain human autonomy, our “decision about who should decide” must be informed by knowledge of how AI would act instead of us; and for AI to be just, we must ensure that the technology—or, more accurately, the people and organisations developing and deploying it—are held accountable in the event of a negative outcome, which would require in turn some understanding of why this outcome arose. More broadly, we must negotiate the terms of the relationship between ourselves and this transformative technology, on grounds that are readily understandable to the proverbial person “on the street”.
Taken together, we argue that these five principles capture the meaning of each of the 47 principles contained in the six high-profile, expert-driven documents, forming an ethical framework within which we offer our recommendations below. This framework of principles is shown in Fig. 2.