Early study: GPT-4 matched or beat humans on theory-of-mind tests; faux-pas misses were guardrails

emollick · x · 2026-08-18

Ethan Mollick resurfaced research showing GPT-4 performed at, sometimes above, human levels across theory-of-mind tests, failing only at detecting faux pas — which turned out to be a guardrail issue rather than a capability gap. In a follow-up reply he also notes a persistent practical issue: advanced LLMs' work products often embed information relevant only to the creator (e.g. traces of earlier drafts), confusing users.

Related event: LLMs match humans on theory of mind yet stumble with multiple perspectives(4 posts)→

Original post →

More from Models

Models channel →