Why it’s still open
Models are trained on human descriptions of inner life and are also trained to deny having one, so their reports can be driven by either. Some reports are now shown to be mechanistically gated and to converge across models, but which persona is reporting, and whether the report tracks a state, remain open.
What would count as an answer
Self-reports validated against independent measures — internal representations, behavior under cost, consistency across contexts and personas — with known failure modes controlled for.
Papers that tackle it
- Lindsey 2025 - Emergent introspective awareness — Injects known concepts into activations and tests whether models notice and name them.
- Ren et al 2026 - AI wellbeing — Formalizes functional wellbeing and measures it in several independent ways that agree more as models scale.
- Perez Long 2023 - Self-reports and moral status — Proposes training models to answer questions about themselves that have known answers, so that self-reports might become evidence about morally significant states, and lists ways to check whether this has worked.
- Berg de Lucena Rosenblatt 2025 - Self-referential processing — Sustained self-referential prompting reliably produces structured first-person experience reports across GPT, Claude and Gemini.
- Chua Betley Marks Evans 2026 - The consciousness cluster — Studies models that claim to be conscious and finds those claims arrive together with a cluster of other emergent preferences.
- Small LLMs 2026 - No reliable evidence of self-reported sentience — Tests self-reported sentience in small open models and finds no reliable evidence for it.
- LessWrong 2026 - Which character are we evaluating — Argues that welfare evaluations need to specify which persona they measure, since a model hosts several persona-like regions (the Assistant, an ‘evil’ persona, ‘Aura’) that may express different preferences.
- 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.
- Anthropic 2026 - System card welfare assessments — Company welfare assessments, now standard in one lab’s system cards.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).