Can a conversational agent pass theory-of-mind tasks? A case study of ChatGPT with the Hinting, False Beliefs, and Strange Stories paradigms.

We investigate the possibility that the recently proposed OpenAI’s ChatGPT conversational agent could be examined with classical theory-of-mind paradigms. We used an indirect speech understanding task, the hinting task, a new text version of a False Belief/False Photographs paradigm, and the Strange Stories paradigm. The hinting task is usually used to assess individuals with autism or schizophrenia by requesting them to infer hidden intentions from short conversations involving two characters. In a first experiment, ChatGPT 3.5 exhibits quite limited performances on the Hinting task when either original scoring or revised rating scales are used. We introduced slightly modified versions of the hinting task in which either cues about the presence of a communicative intention were added or a specific question about the character’s intentions were asked. Only the latter demonstrated enhanced performances. No dissociation between the conditions was found. The Strange Stories were associated with correct performances but we could not be sure that the algorithm had no prior knowledge of the test. In the second experiment, the most recent version of ChatGPT (4-0314) exhibited better performances in the Hinting task, although they did not match the average scores of healthy subjects. In addition, the model could solve first and second order False Beliefs tests but failed on items with reference to a physical property like object visibility or more complex inferences. This work offers an illustration of the possible application of psychological constructs and paradigms to a conversational agent of a radically new nature.