Tom-Felix Thormann

Why do we form moral judgments? One influential answer is that the human capacity for moral judgment is an evolutionary adaptation. P. Kyle Stanford argues that morality is adaptive because it leads individuals to externalize norms, prompting them to evaluate both their own behavior and that of potential partners by a single standard. In this paper, we examine Stanford’s hypothesis using a game-theoretic model. Our results support and extend his proposal, demonstrating that - under the assumptions chosen - externalization is a plausible driver of the evolution of morality as it fosters the learning of cooperation in challenging novel contexts. We find that externalization allows agents to avoid the \textit{exploitation trap}: by resisting unsustainable short-term gains through exploitation, externalizing agents secure the benefits of robust long-term cooperation. By modeling externalization as a constraint on strategy and partner choice, we avoid the common pitfall of game-theoretic approaches to represent morality as a mere disposition toward altruistic behavior.

Human deception involves intentionally causing false beliefs. Since it is contested whether large language models (LLMs) and artificial intelligence (AI) in general have beliefs and intentions, deception by such systems is typically defined in terms of observable behavior. However, this paper argues that such behavioral definitions fall short of the strategic nature of deception, among other things, because they do not consistently distinguish deception from mere error. Moreover, behavioral definitions of AI deception are not optimal as methodological guidance for mechanistic approaches to AI deception because they do not specify candidates for internal markers of deception. It is important to take the internal mechanisms of LLMs and AI systems into account when addressing deception, since skilled forms of deception may be behaviorally indistinguishable from honesty. This paper formulates these challenges to behavioral definitions and addresses them by proposing a novel definition of strategic AI deception in terms of generalized propositional attitudes. Specifically, it provides generalizations of the attitudes of belief and desire that can be ascribed to current and future AI systems less controversially than their human counterparts, yet are expressive enough to figure in a definition of strategic deception. This definition has several advantages over its behavioral alternatives as it 1) better satisfies pre-theoretic intuitions such as the error condition 2) fits better with applied researchers’ usage of the term “strategic deception”, 3) is relevant to large-scale risks of AI deception, and 4) is methodologically instructive for mechanistic approaches to deception.

According to an increasingly influential view, the complex, human-like behavior of large language models (LLMs) should be explained as roleplay or simulation of characters. Such notions aim at enabling intentional explanations without ascribing mental states to LLMs. However, there are a multitude of roleplay views that are applied inconsistently in empirical explanations of LLM behavior. Moreover, it is unclear what exactly different roleplay views predict and, consequently, how they can be empirically distinguished from each other and from explanations that do ascribe mental states to LLMs. In this paper, we provide a taxonomy of roleplay views structured through three dimensions: Interpretation (how can LLM behavior be best explained according to the view), Mechanism (what commitments does the view have about internal mechanisms of roleplay), and Generality (does the view aim to explain all LLM behavior or specific kinds). We then analyze, with regard to the case of LLM deception, why it is practically relevant whether LLMs are mere roleplayers or in possession of genuine mental states. However, we identify two challenges for empirically distinguishing these two: Similarity (Roleplay is inherently difficult to distinguish from intentional behavior) and Ambiguity (it is often ambiguous what roleplay views predict). Finally, we propose three strategies for overcoming these challenges, making progress towards transforming roleplay from loose metaphor to a testable theory.