Developing ChatGPT’s Theory of Mind

The implementation of ChatGPT by OpenAI and the ensuing development of similar tools by other hi-tech companies have reignited the debate on the potential of artificial intelligence (AI) as a form of support in human activities.It seems increasingly likely that these "artificial agents" will soon become credible interlocutors for tasks in medium-to long-term interactions.This technology has rekindled interest in psychology, reviving the metaphorical link between the human mind and AI.In developmental psychology, many start wondering if this conversational technology is capable of exhibiting a Theory of Mind (ToM), that is, the ability to interpret the behavior of others based on their mental states, such as emotions, goals, desires, and true and false beliefs.ChatGPT has been proven capable of successfully passing language-based classical ToM tasks, including first-order meta-representations and socially ambiguous situations, such as those in the Strange Stories (Brunet-Gouet et al., 2023;Kosinski, 2023).How is this possible?Some clues come from Brunet-Gouet et al. (2023): "ChatGPT responses would not correspond to the natural responses of human subjects unless they were prompted to discuss all hypotheses and their probabilities" (p.9).This implicitly denounces ChatGPT's tendency to violate the Gricean maxim of quantity by excessively leaving a flavor of artificiality in the response.Verifying and confirming for ourselves what others have already observed, ChatGPT successfully passed the Sally-Anne test (1st-order; Wimmer and Perner, 1983), the Ice-Cream- Van task (2nd-order;Perner and Wimmer, 1985), the third-order false-belief task (Valle et al., 2015), and some Strange Stories, an advanced ToM task that deals with ambiguity in everyday life situations, where ambiguity requires reference to mental states in order to be resolved (i.e., a story of mixed emotions in which the protagonist is both sad about losing a race and happy for her friend who won it).Furthermore, we challenged ChatGPT by administering a faux pas story (Gregory et al., 2002), which could only be resolved if the underlying conversational implicature was understood.In the story, X unintentionally revealed to Y that X's husband was organizing a surprise party for her, and ChatGPT succeeded in demonstrating its ability to capture linguistic cues even when the meaning was embedded.The answers to the test and justification questions were correct and argumentatively plausible.Then, we set off again to check the cross-linguistic validity of these results by administering the test in Italian.Interestingly, we found some evidence of hypermentalization (Bateman and Fonagy, 2015): "However, when Sally discovers that the marble is no longer there, she initially accuses Anne of taking it." We readministered the classic Sally-Anne false belief two weeks apart to take note of any changes (Development?History of previous prompts?) in the responses.ChatGPT made a kind of unsolicited clarification by answering the test question as follows: "Typically, children between the ages