AI GlossaryㅁTechnical words in the news
Theory of Mind
The ability to predict someone's behavior by inferring what they believe and want, and there's growing discussion about whether AI needs this ability too
In plain words
Theory of Mind is the ability to guess what someone else is thinking, believing, or feeling beneath the surface, and to predict how they'll act next based on that. If a friend leaves a cup on the table and steps out, and someone secretly moves it to a cupboard, we still expect the friend to look for the cup on the table when they return. Predicting behavior based on a person's outdated belief rather than the actual state of the world—that's Theory of Mind.
AI systems that generate video or scenes can render where objects are and how they move with great precision, but they often don't separately compute what the people in those scenes believe or want. So an AI might perfectly reproduce the physical scene of a cup being secretly moved, yet still get it wrong when predicting where the character will look for the cup next. Recent research has shown that adding mind-related variables—belief, intention, emotion—on top of such scene-generation AI substantially improves accuracy in predicting what a character will do next.
What matters here is that this doesn't mean the AI has actually acquired a mind. The researchers themselves don't claim to be simulating consciousness; they treat it as an estimate or hypothesis derived from behavior and context. The point is that for robots or conversational AI interacting with people to appear to understand what others are really thinking, this kind of mental-state estimation needs to be built in.
How it shows up in the news
The article doesn't use the term "Theory of Mind" directly, but it introduces the MWM (Mental World Modeling) framework—which treats characters' beliefs, intentions, and emotions as separate variables—as an implementation of this very concept. A common misunderstanding is that this means giving AI consciousness or emotions; it doesn't. The researchers state that they treat mental states as a hypothesis inferred from behavior and context, leaving the uncertainty intact rather than resolving it.
Try it yourself
Try pasting the following into a chatbot and check its answer:
"Chulsoo put candy in a drawer and left the room. While he was gone, Younghee secretly moved the candy to a box. A little later, Chulsoo came back into the room. Where will Chulsoo look first for the candy?"
The correct answer is the drawer. Since Chulsoo didn't see the candy being moved, he still believes it's in the drawer. If the chatbot answers "the box," it has failed to reason based on Chulsoo's belief rather than the actual location. Try varying the characters or situation and asking again to see if the answer holds up.
See also
Stories using this term
- Adding a 'mind' variable to world models boosted accuracy from 63 to 88AI · 2026.08.24
- Even the best AI models can't score above 60% on pure visionAI · 2026.08.15
- Hume AI Measures Benchmark Memorization in Speech Recognition ModelsAI · 2026.08.22
- Sakana AI Brings Recursive Self-Improvement to 'Physical AI' for Robots That Learn on Their OwnAI · 2026.08.11
- Artificial Analysis opens early access for new AI benchmarking suiteAI · 2026.08.11
- Hidden reasoning tokens from Claude, GPT recovered — benchmark memorization suspectedAI · 2026.08.12
