Luca Alberto Rappuoli

The recent rise in popularity of large language models (LLMs) has prompted considerable concerns about their moral capabilities. Although considerable effort has been dedicated to aligning LLMs with human moral values, existing benchmarks and evaluations remain largely superficial, typically measuring alignment based on final ethical verdicts rather than explicit moral reasoning. In response, this paper aims to advance the investigation of LLMs’ moral capabilities by examining their capacity to function as Artificial Moral Assistants (AMAs), systems envisioned in the philosophical literature to support human moral deliberation. We assert that qualifying as an AMA requires more than what state-of-the-art alignment techniques aim to achieve: not only must AMAs be able to discern ethically problematic situations, they should also be able to actively reason about them, navigating between conflicting values outside of those embedded in the alignment phase. Building on existing philosophical literature, we begin by designing a new formal framework of the specific kind of behaviour an AMA should exhibit, individuating key qualities such as deductive and abductive moral reasoning. Drawing on this theoretical framework, we develop a benchmark to test these qualities and evaluate popular open LLMs against it. Our results reveal considerable variability across models and highlight persistent shortcomings, particularly regarding abductive moral reasoning. Our work connects theoretical philosophy with practical AI evaluation while also emphasising the need for dedicated strategies to explicitly enhance moral reasoning capabilities in LLMs.

According to McDowell (1979), Wittgenstein’s considerations on rule-following pave the way for accepting the idea that our moral outlook is essentially irreducible to any set of rules of action. For every finite set of moral rules, McDowell maintains, there will always be a situation in which the mechanical application of those rules would yield an action that a perfectly ethical agent would not have performed in the same circumstances. In this paper, I argue that embracing this thesis would commit us to the demanding idea that there can never be an AI system that could properly be described as ethical. Owing to its algorithmic nature, AI is constitutively incapable of behaving in any way that resists description in terms of acting in the light of a formulable rule. It follows that, if the moral outlook of an ethical agent is constitutively uncodifiable, there is no possible set of rules on which an AI system could rely to render its actions genuinely akin to those of an ethical agent. Yet Wittgenstein’s reflections, I contend, do not warrant so far-reaching a conclusion. His remarks on rule-following should instead be read as revealing the irreducibility of moral understanding to propositional knowledge – its epistemic opacity – rather than as establishing its metaphysical uncodifiability.

Politicians select their words meticulously, never losing sight of their ultimate communicative goal. Sometimes, their objective may be that of not being fully understood by a large portion of the audience. They can achieve this by means of dogwhistles; linguistic expressions that, in addition to their literal meaning, convey a concealed message to a specific sub-group of the audience. This paper focuses on the distinction between overt and covert dogwhistles introduced by J. Saul (2018). I argue that, even if the distinction successfully captures a genuine divide within the category of dogwhistles, the account proposed by Saul to explain the distinction is unsatisfactory. In response to this state of affairs, I illustrate how the distinction between overt and covert dogwhistle can be refined and illuminated by incorporating it into the 'Simple Theory' of dogwhistles advanced by J. Khoo (2017).