Theory of Mind: What It Means in Psychology and Why AI Researchers Are Debating It

What theory of mind means in psychology

Theory of mind (ToM) is defined as the capacity to understand other individuals by attributing mental states to them, according to Wikipedia's overview of the concept. It lets a person recognize that someone else may hold beliefs, desires or intentions that differ from their own, and use that recognition to predict or explain behavior.

The term entered modern research with the 1978 paper 'Does the chimpanzee have a theory of mind?' by David Premack and Guy Woodruff, which pushed psychologists to study the ability across species and across childhood development, per Wikipedia. Two rival explanations later emerged for how ToM works: 'theory-theory', which holds that people apply folk-psychological rules to infer others' states, and simulation theory, which holds that people mentally model others by imagining themselves in their place.

The Sally-Anne test

The standard tool for testing ToM in children is the false-belief task, most often presented as the Sally-Anne story: Sally puts a marble in her basket and leaves the room; Anne then moves the marble to a box. A child passes the test by predicting that Sally will look for the marble in the basket — her original, now false, belief — rather than where it actually is, Wikipedia explains. A related version, the unexpected-contents task, asks children what someone else will think is inside a box that is labeled misleadingly.

Children typically develop these abilities in a set sequence between roughly ages three and five, moving from understanding that people can want different things, to understanding that people can hold different beliefs, to passing full false-belief tests. Wikipedia notes that the exact sequence varies by culture: children in Iran and China, for example, grasp 'knowledge access' earlier but take longer to grasp 'diverse beliefs', which researchers link to collectivist cultural values.

Why artificial intelligence entered the debate

Interest in machine theory of mind grew after Stanford computational psychologist Michal Kosinski ran the same false-belief tasks used on children through a series of language models. According to the working-paper page on Stanford Graduate School of Business's own site, Kosinski built 40 false-belief scenarios, plus matching true-belief control versions, and tested them on language models released between the late 2010s and 2023.

A separate write-up of the research on PsyPost gives different figures for what it describes as a later round of testing covering 11 models: it reports GPT-3 variants solving only about 20% of tasks, and 'ChatGPT-4' solving 75% overall, with a split of 90% on unexpected-contents tasks but only 60% on the harder unexpected-transfer version. The two sources describe different test batches and model versions, which is one likely reason the percentages do not match; both, however, point to a clear improvement in newer models compared with older ones.

The pushback: pattern matching or genuine understanding?

Kosinski's conclusion has not gone unchallenged. Harvard psychologist Tomer Ullman published a paper on arXiv titled 'Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks', arguing that small rewordings which preserve the same underlying psychological logic can flip a model's answer from correct to incorrect. Ullman's position, summarized in the paper's own framing, is that researchers should start from a skeptical baseline when evaluating machine psychology, and that scattered failures on near-identical tasks matter more than an average success rate.

PsyPost's coverage of the Stanford research also flags this limitation directly: because GPT models are trained on enormous amounts of text that may include similar false-belief stories, it is difficult to rule out that a model is recognizing a familiar pattern rather than reasoning about an unseen character's mental state. The researchers attempted to design new scenarios to reduce this risk, but the article notes this cannot be fully guaranteed.

Why the question matters beyond the lab

The same test battery used on children and chimpanzees is now being applied to software because, as the Stanford GSB page notes, theory of mind is central to human social interaction, communication, empathy, self-consciousness and morality. Both the Stanford material and the Ullman paper agree, however, that passing a fixed set of test questions is not the same as establishing that a system possesses theory of mind in the way cognitive scientists use the term for humans.