Why it’s still open
There is now real but limited evidence of introspective access, from concept injection and self-prediction. It is narrow, unreliable and hard to separate from learned knowledge about models in general, and philosophers disagree on when it deserves the name introspection.
What would count as an answer
Self-reports that track internal states with measurable accuracy against a ground truth, generalize beyond the trained cases, and cannot be explained by third-person knowledge the model could have learned from text.
Papers that tackle it
- Lindsey 2025 - Emergent introspective awareness — Injects known concepts into activations and tests whether models notice and name them.
- Betley et al 2025 - Tell me about yourself — Models fine-tuned on implicit behaviors (risky choices, insecure code) can describe those behaviors without being trained to, and can attribute different learned policies to different personas.
- Binder et al 2025 - Looking inward — Models trained to predict their own behavior do better at it than other models trained on the same data, which is evidence of privileged self-access on simple tasks.
- Comsa Shanahan 2025 - Introspection in LLMs — Examines under what conditions an LLM’s statements about itself deserve to be called introspection, working through example cases.
- Plunkett et al 2025 - Self-interpretability — Shows that models can describe features of the internal processes behind their own decisions, and that training improves this ability.
- Kadavath et al 2022 - Language models mostly know what they know — Finds that large models are fairly well calibrated about whether their answers are right and can be trained to predict whether they know an answer.
- Macar et al 2026 - Mechanisms of introspective awareness — The circuit behind injected-thought detection, and where it comes from in training.
- Singh Linzen Ravfogel 2026 - Can LLMs introspect — A skeptical check: current tests don’t separate introspection from cues in the input.
- vgel et al 2026 - Latent introspection — Self-reports may understate internal access.
- Hahami et al 2025 - Detecting the disturbance — Removes a confound from injection tests.
- Fraser-Taliente et al 2026 - Natural language autoencoders — Reading unverbalized beliefs out of activations.
- Shenoy et al 2026 - Introspection adapters — Training models to report what they learned.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
- Which activations belong to the Assistant
- Is there a global workspace
- When self-report counts as welfare evidence
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).