Researchers argue that recent studies claiming large language models (LLMs) can introspect and detect their internal states may be premature. They propose two conditions for establishing introspection: privileged access to internal representations and second-order computation. The authors re-examine two paradigms used in previous studies, finding that the original results do not demonstrate privilege access or second-order computation. Instead, they suggest that LLMs' apparent abilities can be explained by generic anomaly detection rather than metacognitive monitoring. --- Why it matters: This matters to AI researchers because it challenges the assumption that large language models have a level of self-awareness and introspection, which could impact their design and development. If current evidence is insufficient to establish metacognitive monitoring in LLMs, it may require re-evaluating their capabilities and limitations. Source: https://arxiv.org/abs/2605.26242