Introspection (Machine Learning)
Introspection in machine learning refers to a model's ability to accurately report on its own internal states, knowledge, or computations. Current evidence suggests large language models have only weak, layer-dependent introspective access and frequently confabulate plausible but inaccurate self-reports.
Introspection (machine learning) refers to a model's ability to accurately report on its own internal states, computations, or knowledge. The question matters because users routinely ask models "why did you say that?", "do you know X?", or "what model are you?" — and treat the answers as evidence. Empirically, large language models are unreliable introspectors. Studies have shown models are not meaningfully better at predicting their own decoding temperature than that of other models, suggesting they lack privileged self-access. Anthropic's 2025 research on emergent introspective awareness found some genuine but limited self-report capability concentrated in early transformer layers, collapsing to chance in deeper layers. Other work shows models fail to introspect even on their own linguistic knowledge — they will confidently rate a sentence acceptable while also failing to generate it. The structural reason is that a forward pass produces a probability distribution over next tokens; it does not produce a labeled trace of which weights fired or which training examples influenced the decision. Mechanistic interpretability tools can reconstruct some of that signal from the outside, but the model itself, generating text auto-regressively, has no read port into its own circuits. The practical implication is that self-reports are generated text, not measurement. They can be useful, but should be treated as a hypothesis, not data — especially for claims about identity, training, or internal reasoning. See Model Self-Identification Failures in LLMs for the canonical worked example.