Can LLMs Introspect? A Reality Check

A new paper argues that current tests do not prove LLMs can truly introspect.

Models often appear self-aware by using anomaly detection, task cues, or pattern matching rather than privileged access to hidden states.

When stronger controls are added, performance drops toward chance, weakening claims of genuine metacognition.

The findings matter because they affect how researchers interpret model self-reports, reliability, and future cognition evaluations.