Beta-testing the ethics plugin

What I mean by ‘radical questions’ (‘radical’ meaning ‘root’), are those that articulate dilemmas that call into question even our ability to answer by pointing to examples, exemplars, or everyday standards. The undermining of this ability, by calling into question those standards, is what makes a question radical. Ethical questions are often prime instances, though not the only ones. Consider the question of whether or not some particular person, thing, or act is good. This cannot always be answered by pointing to another person, or this or that thing or act. Such cases must be answered at least in part by considering the further question ‘What is Good?’, otherwise we would be left attempting to compare particulars without a basis of comparison. Such radical questions are readily found in ancient Greek philosophy, especially in Plato and Aristotle, where they are called aporiai. These aporiai are amenable only to answers that are general, unitary, and explanatory standards for judgement. I am indebted to Vasilis Politis for his notion of radical aporia. For more on this topic, I recommend especially his two recent monographs on Plato.

The analytical tradition itself is not much older than computer science and, at root, both arose from the same advances in formal logic and mathematics at the end of the nineteenth century. It is rarely mentioned that the tripartite division was invented during the twentieth century and may often be misleading. Anscombe coined the term ‘consequentialism’ in her article ‘Modern Moral Philosophy’ from 1958, which was also a locus for virtue ethics. The term ‘deontology’ was first introduced by Bentham, in his Chrestomathia from 1816, as a synonym for ‘Dicastic Ethics’, addressed to the will, as opposed to the ‘Exegetic’ or ‘Expository’, addressed to the understanding. The term was reintroduced by C. D. Broad, in his Five Types of Ethical Theory from 1930, and used in reference to types of action considered regardless of their consequences. I say all this not in simple-minded veneration of the past or aversion to recency, but merely to point out that beginning from such a palette restricts our view, has the potential to beg important questions, and sometimes commits the reverse of those fallacies.

Take utilitarianism, which, although recognising agents, is agent-neural. That is, it cannot be combined with a non-agent-neutral theory. It follows from such incompatibilities that, whatever the ADC approach is, it is strictly speaking not a combination of the three normative approaches. A full investigation would be beyond the scope of the present short article. However, it should be clear to see that this objection can be almost symmetrically extended to (at least) the other two main kinds of normative theory.

Wernaart (2021, 9) notes that “they move away from the debate on what a machine should do, and instead focus on what we want it to do” (My emphasis added). This is, I believe, a misunderstanding of the situation caused by the notion that because one is avoiding explicitly using the normative palette, one is doing non-normative ethics.

Aliman and Kester say that “Instead of specifying what an agent ought to do, AU helps to identify what the current society should want an (artificial or human) agent to do if this society wants to maximize expected utility” (Aliman and Kester 2019, 4, as cited in Wernaart 2021, n. 92). There are clearly two normative elements already involved here. The first is that AU is intended to help to identify what should happen. The second is that it relies upon prior normative conceptions of utility.

The authors of this approach do not address this well-known fallacy of the derivation of an ought from an is, which goes back to Hume (Treatise III.1.1). In light of this, their approach would, at the very least, require a sustained attempt at philosophical justification in that regard. For example, Steven Kraaijeveld is more careful in noting that although empirical findings may be useful for informing some areas of ethical enquiry, the role is a supportive one that does not generate normative conclusions in the absence of prior normative premisses.

This is wisely recognised by the authors of the Moral Machine experiment, which is “a multilingual online ‘serious game’ for collecting large-scale data on how citizens would want autonomous vehicles to solve moral dilemmas in the context of unavoidable accidents” (Awad et al. 2018, 59, as cited in Wernaart 2021). When reflecting on their study, Bonnefon said that “Considering our results to be normative would amount to saying that there is no objective definition of what is moral or immoral; morality would therefore be a social construction, limited in space and time. Consequently, morality would be whatever the population thinks is moral, here and now. Moral Machine would therefore be the arbiter and standard of morality. Personally, I think this interpretation is insane, but the philosophical debate is beyond me. All that my coauthors and I could do was say that we never had any such ambitions!” (Bonnefon 2021, p. 132).

The palette is also mentioned in this context due to the salience in the debate of these so-called “classical” normative frameworks; see footnote 2. There is a potential here that the framework itself will rule out certain views, but this is not determinable on the present characterizations that are available.

I am grateful to an anonymous reviewer for pushing me to clarify this point, especially in view of the fact that some researchers may be unaware of it.

It could be objected that, because we are concerned with ethical issues, we as matter of course need to appeal to prior ethical theories. However, ethical theories are the results of prior philosophical investigations, not the beginning of them. Such objections would not pass muster in historical contexts in which there were no explicit prior theories, or they were not named, etc.

Thereby, the ancient aporiai re-emerge. See footnote 1 on radical questions.