AI & Theory of Mind

Key Terms
Theory of Mind – A hallmark of social intelligence that allows us to infer the inner reality of another person’s mind based on their speech and behaviour, as well as our own knowledge of human nature.

False-belief test – A common experiment to see if someone hasa theory of mind, checking whether they understand others can hold beliefs that are wrong.

Scaling laws – In AI or biology, patterns showing how performance or behaviour changes predictably as size, data, or resources increase.

Anthropomorphism – Attributing human characteristics or emotions to non-human things.


Main Text
In psychology, Theory of Mind (ToM) is the ability to infer others’ mental states—beliefs, desires, intentions—and recognise they may differ from one’s own. It is essential for social reasoning, moral judgement, and self-awareness. Developmentally, children usually acquire ToM around age four or five, often demonstrated through “false-belief” tests such as the Sally-Anne scenario, where they correctly predict that a character will act based on their mistaken belief rather than reality.

In recent work, researchers have investigated the degree to which large language models (LLMs) like GPT-3.5 and GPT-4 possess ToM-like abilities. In some studies, these models performed well on false-belief tasks, achieving accuracy comparable to a seven-year-old. However, small modifications in the scenario caused sharp drops in performance, showing a fragility unlike the human ToM. With no embodiment and no real-world experience, LLMs are reliant on linguistic patterns, meaning their apparent understanding differs fundamentally from the human version.

Misrepresenting concepts such as ToM in AI discussions risks cultivating anthropomorphism and exaggerating abilities. In the article, experts note that effective responses from LLMs can mask the appearance of true understanding, much like early chatbots fooled users despite their limited ability. Avoiding such confusion requires careful definitions, acknowledgement of limitations, and clear distinctions between imitation and true cognitive ability.