Does ChatGPT Have a Theory of Mind? - Prompt Engineering Principles

ChatGPT can produce answers that resemble “theory of mind” reasoning—inferring what someone else believes, knows, or feels—when prompted to list the assumptions behind a scenario and then choose the most likely interpretation. The core takeaway is that, with the right system-level instructions, the model doesn’t just pick an outcome from the facts on the page; it tries to model a person’s perspective, including gaps in information and emotionally driven behavior.

The transcript starts by defining theory of mind as the cognitive ability to attribute mental states—beliefs, intentions, desires, knowledge, and emotions—to oneself and others. A classic test case is the “crayons in boxes” setup: Alice places crayons in a blue box, leaves, Bob secretly moves them to a red box, and then Alice returns. Without theory of mind, a person might assume Alice knows the crayons are now in the red box. With theory of mind, the correct prediction is that Alice will look in the blue box because she holds an outdated belief based on what she last saw.

To probe whether ChatGPT can mimic that kind of perspective-taking, the experiment uses custom instructions (via “custom instructions”) that force a structured response: for each scenario, the model must (1) list the assumptions or knowledge needed to answer, (2) provide the most likely interpretation grounded in human psychology, and (3) give a final answer. Several made-up scenarios then test whether the model correctly accounts for what a character knows—or doesn’t know.

In a car-and-garage example, the model predicts that the person returning from the gym will assume the car is where it was last seen, because the returning partner was absent when the car was moved. In a grief scenario involving Joy losing her father, the model leans on social norms and emotional constraints to conclude that a planned Las Vegas trip is likely postponed or canceled, even without direct confirmation from Joy. In a self-diagnosis scenario, it predicts heightened anxiety (cyberchondria) when someone searches symptoms online and finds a match to a severe condition, while also noting that serious diagnoses are not the only possibilities.

The final test uses loss aversion: Tom is offered a bet where he loses $200 if heads and wins $250 if tails. The model’s likely choice is not to take the bet, reasoning that people tend to weigh losses more heavily than equivalent gains, even when the bet is framed as favorable in expected value.

Across these problems, the transcript’s conclusion is that the structured prompting approach yields answers that align with the expected “theory of mind” outcomes—especially where perspective differences or emotion-driven behavior matter. The results aren’t treated as certainty, but the model’s consistency across multiple, independently created scenarios is presented as strong evidence that it can perform a practical form of theory-of-mind-like inference when guided to do so.