In recent years, the capabilities of Large Language Models (LLMs), such as ChatGPT, to imitate human behavioral patterns have been attracting growing interest from experimental psychology. Although ChatGPT can successfully generate accurate theoretical and inferential information in several fields, its ability to exhibit a Theory of Mind (ToM) is a topic of debate and interest in literature. Impairments in ToM are considered responsible for social difficulties in many clinical conditions, such as Autism Spectrum Disorder (ASD). Some studies showed that ChatGPT can successfully pass classical ToM tasks, however, the response style used by LLMs to solve advanced ToM tasks, comparing their abilities with those of typical development (TD) individuals and clinical populations, has not been explored. In this preliminary study, we administered the Advanced ToM Test and the Emotion Attribution Task to ChatGPT 3.5 and ChatGPT-4 and compared their responses with those of an ASD and TD group. Our results showed that the two LLMs had higher accuracy in understanding mental states, although ChatGPT-3.5 failed with more complex mental states. In understanding emotional states, ChatGPT-3.5 performed significantly worse than TDs but did not differ from ASDs, showing difficulty with negative emotions. ChatGPT-4 achieved higher accuracy, but difficulties with recognizing sadness and anger persisted. The style adopted by both LLMs appeared verbose, and repetitive, tending to violate Grice's maxims. This conversational style seems similar to that adopted by high-functioning ASDs. Clinical implications and potential applications are discussed.
Read full abstract