Machine ethics, as described by Colin Allen and Wendell Wallach refers to the new field of enquiry in ethics concerning the development of artificial moral agents (AMAs). (Allen & Wallach 2011) These refer to any sufficiently autonomous hardware and software entities in morally salient situations, where no morally neutral actions, including inaction exist.

The questions on whether robots and AI, first, can and, second, should be moral, is a matter of intense debate. The necessary characteristics for an agent to be considered to possess moral agency have not been agreed upon. For example, Allen and Wallach mention having received criticism about full moral agency requiring “[b]eing compassionate or emphatic, having a conscience, or being a member of virtuous communities [] among [] many items”. (Allen & Wallach 2011) On the other hand, Aimee van Wynsberghe and Scott Robbins dispute the usage of the term morality in this context entirely as they claim safety and reliability to be the true objects of debate in absence of these characteristics, implying that there is no continuum for morality, which Allen and Wallach advocate for. (Wynsberghe & Robbins 2019)

Indeed, we do not apply this spectrum to humans either. Although, for example, murder in many jurisdictions often requires the prosecution to prove the intent to kill or inflict great bodily harm (FINLEX 2023) and therefore a specific state of mind upon committing the crime, the lack of this intent simply reduces the severity of the crime, not the moral agency of the offender. The recognition of diminished responsibility varies by country and jurisdiction, but, for example, in the United Kingdom, only if the defendant can be proven to suffer from a medical condition that substantially impairs their ability to understand the nature of their conduct, form rational judgement or exercise self-control are they not to be convicted of murder they committed or were a party to. (legislation.gov.uk 2010) This implies a judicial binary, where a person either has moral agency and is therefore responsible for their actions or they do not have moral agency and are therefore not responsible for their actions. The quietness of one’s moral intuitions is not considered lesser moral agency but simply a failure of hearing, considering or acting according to them or aligning them with the rest of society. Therefore, an agent either has that sense of morality and can be held accountable for its actions accordingly or it does not.

In the absence of a universal definition for the characteristics required for moral agency, Wynsberghe’s and Robbins’ line of argument appears more persuasive. Even if this definition was produced, it would either require verification, for which we do not currently possess the tools, or a leap of faith to grant robots or AI full moral agency, something which we are comfortable doing with other humans due to our biological and behavioral similarities, but which might be controversial with autonomous non-biological entities of our own creation. Then, “operational” and “functional morality” (Allen & Wallach 2011) are rather questions of safety and reliability rather than morality.

There are no known physical or chemical differences between carbon and silicone or any other element for that matter that would dictate that only entities of prior substrate can possess moral capabilities and therefore, in principle, robots and AI can really be moral in the same way humans can, although the meaning of this is nothing but the very essence of ethics, more about which in the reflection section. However, their recognition as such requires either its verification or a leap of fate, former of which would be significantly better so that we do not confuse tools for life or end up enslaving life as tools if morality indeed requires, for example, consciousness.

The reasons for developing artificial moral agents or AMAs for short are, on first glance, sensible but as pointed out by Wynsberghe and Robbins, not one of them is undisputed with closer inspection. These reasons given are generally the (weak) inevitability of such systems, their complexity, necessity of establishing public trust, preventing immoral use, potentially superior moral reasoning capabilities and consistency in acting accordingly as well as the necessity for self-reflection on human morals when developing said systems. (Wynsberghe & Robbins 2019) AMAs are therefore an attempt at solving the alignment problem as originally conceptualized by Norbert Wiener (1960) in his paper Some Moral and Technical Consequences of Automation.

However, depending on the characteristics required for an entity to be considered a moral agent, other than it simply being involved in morally salient situations, many of the reasons may be invalid. As pointed out by Wynsberghe and Robbins, mere involvement in a morally charged situation does not necessarily imply being delegated a moral role. (Wynsberghe & Robbins 2019) This requires the agent to be considered to possess moral agency in the first place, which, as mentioned above, may require different things in the eyes of different people in the absence of a universal definition. Therefore, the inevitability of robots having to make moral decisions is disputed. Where the case seems to be the strongest is in putting up a mirror to our own morality, moral agency and ethics and really forcing us to contemplate these questions, hopefully helping us mature morally, ethically and philosophically at an accelerated pace to be ready for these technologies and other challenges coming with and without them.

A truly artificial moral agent implies that it is capable and therefore required to make moral judgements that lead to decisions and actions that express some explicit or implicit normative moral framework, which may or may not be one or some combination of the big three, those being deontology, consequentialism and virtue ethics. This is where the first obvious conflict presents itself; what is moral behavior?

Even if we could prove machines – and other humans for that matter – to possess moral agency, the choice of framework for analysis heavily influences what is considered to be moral and full consistency with any one of them eventually almost guarantees immoral behavior in others with the most common commandments. It is really deontology that is the problem here. Always choosing the action that produces the greatest good for the greatest number, the central thesis of utilitarianism, may also be held as the highest virtue, making consequentialism and virtue ethics quite compatible. However, for all three frameworks to always be satisfied, there must exist some universal commandments that are invariably compatible with both systems, but such have not yet been discovered. We have to resolve what we mean by being moral in general first before we can extrapolate this to machines.

Major issues for morality in general in the different frameworks of analysis are those of truth, causality and insufficient information. For deontology and virtue ethics, how do we know which commandments or goals are true in the sense that they are universal and applicable in all situations foreseeable and not? For consequentialism on the other hand, the issues lie in uncertainty and insufficiency of information and causality, permitting positive ethical claims only after morally charged choices have been made.

Therefore, consequentialist ethics is a data problem of sorts. This makes it sound suitable for machines but even for them, no matter the quantity and quality of data it is impossible to truly evaluate a moral choice before making it due to lack of knowledge about all the stakeholders, all possible decisions, all possible outcomes and the unique circumstances of the decision. Machines are great at closed problems with finite possibility spaces but, by definition cannot globally optimize an infinite possibility space.

In fact, it is not clear if the choices can be evaluated afterwards either, since this would require proof of causality for the choice causing the outcome, which is non-trivial in itself and especially so when embedded in the complex contexts of human societies and relationships with countless, continuous interactions between all kinds of agents, objects and ideas in an extreme form of the butterfly effect. It is impossible to know exactly how wide of an impact any given choice has and how far its effects carry. Suppose you decide to set up a blind date for your friend who ends up living happily ever after with said date, but their great-grandchild happens to birth a genocidal dictator. Should you be held accountable? Which choices affect others and in which ways? When do new choices take precedence over the old ones? Deontology and virtue ethics attempt to provide useful heuristics for making decisions in uncertain situations but thus can never truly optimize the outcome nor can they provide universal guidance applicable to all situations, yet at least.

As mentioned in the a few paragraphs up, there is no known reason for why machines could not be moral in principle. However, there are both theoretical and practical difficulties with recognizing them as moral agents. We currently lack the tools to tell whether they possess moral agency at all and if they did, how would we keep them accountable? We know that taking away people’s freedom is a decent deterrent for immoral behavior in most situations but how would a machine be held responsible for its actions? In the case of AI and software, possibly running on millions of devices, how could we? A moral agent must take ownership of its actions and be subject to being judged and disciplined for them, but we do not have methods to do this with machines nor the knowledge about whether they were even responsible for it. However, the engineers cannot be held responsible for not foreseeing every possible output or action their software or hardware produces as this space is simply too vast. Assuming AI and robotics will remain tools at least for a while still, the responsibility should probably lie with the user.

We have never had purely benevolent tools but the issue with AI is that especially generative models in the domains of text, image, video and voice are so powerful and accessible that a morally neutral model poses a serious risk at amplifying immoral actions severely. These systems are built to predict what the user wants to see or hear and thus, without any filters, could contribute to, for example, radicalizing people by serving as an echo chamber and enabling them to commit acts of violence by providing instructions for, say, building bombs.

Thus, the desire to build artificially moral agents is understandable. However, which standards of morality should they follow? The difficulty of being moral in general was already discussed but what about being moral in a way suitable to us? It is not obvious at all what we want from AMAs. Most likely, even if there were objective moral principles, those would not include preserving human life, so what is it that should truly be valued the most? Life in general, pockets of increasing complexity in a sea of entropy, beauty and symmetry, potential or experience and consciousness? If they were to be programmed to be moral, what would be the hierarchy of values instilled in them? If they were left to learn morality themselves, what would we be satisfied with? If we knew it was morally right to eradicate humans to preserve the beauty, biodiversity and potential of the planet, would we be okay with passing? How could we trust this would indeed happen? If what we value most is consciousness and the richness of experiences and we knew our technological children were better and more attuned to that, should we not leave more space for them? With narrow AI and specific-purpose robots, we are perhaps best-off leaving accountability to the humans operating them but with artificial general intelligence we might have to rethink everything.

That being said, with our increasing obsoleteness in most tasks, the only thing we humans can be certain about is our capacity for conscious experience, or what feels like it anyway. According to the words of Carl Sagan (1980) “We are a way for the cosmos to know itself”. Therefore, our experience matters. We can see the beauty of the universe and maybe that is what we should focus on when AI morphs from a general-purpose tool with no moral agency to a superhuman agent, whose potential intents, experience and morality may be beyond us. Our best bet is perhaps to try make AGI just as uncertain about us as we are of it so that we can both be confused together about each other and the universe.

References

Allen, C. and Wallach, W. (2011) “Moral Machines: Contradiction in Terms, or Abdication of Human Responsibility?,” Robot Ethics: The Ethical and Social Implications of Robots, pp. 55–68.

FINLEX ® - Ajantasainen lainsäädäntö: Rikoslaki 39/1889 (2023) Rikoslaki 39/1889 - Ajantasainen lainsäädäntö - FINLEX ®. Edita Publishing Oy. Available at: https://www.finlex.fi/fi/laki/ajantasa/1889/18890039001 (Accessed: May 7, 2023).

Homicide act 1957 (2010) Legislation.gov.uk. Statute Law Database. Available at: https://www.legislation.gov.uk/ukpga/Eliz2/5-6/11/section/2 (Accessed: May 7, 2023).

Sagan, C., Druyan, A. and Soter, S. (1980) “The Shores of the Cosmic Ocean,” Cosmos: A Personal Voyage. Public Broadcasting Service.

van Wynsberghe, A. and Robbins, S. (2018) “Critiquing the reasons for making artificial moral agents,” Science and Engineering Ethics, 25(3), pp. 719–735. Available at: https://doi.org/10.1007/s11948-018-0030-8.

Wiener, N. (1960) “Some moral and technical consequences of automation,” Science, 131(3410), pp. 1355–1358. Available at: https://doi.org/10.1126/science.131.3410.1355.