Modelling what someone else knows, wants or believes, including when their beliefs are wrong. Language models pass many of the standard tests for this, and there is a live argument about whether that reflects a genuine capability or familiarity with the test formats from training data.
Technical view
Assessed with false-belief tasks and variants. Frontier models perform well on classic formulations and degrade sharply on perturbed versions designed to break surface pattern matching, which is the crux of the dispute over whether the capability is real. It matters practically for assistants that must track user knowledge, and it is a prerequisite capability for the kinds of manipulation and deception that alignment research is concerned with.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.