Emergent Introspective Awareness in Large Language Models is an Anthropic interpretability paper by Jack Lindsey and Anthropic collaborators, published in 2025 at transformer-circuits.pub/2025/introspection. It investigates whether large language models have a functional ability to introspect on their own internal states, as distinct from generating plausible-sounding self-descriptions, and reports that Claude Opus 4 and 4.1 show the greatest such ability among the models tested, while characterizing that ability as highly unreliable and context-dependent.
Summary of argument
The paper asks whether introspective self-reports by language models reflect genuine examination of internal states or are confabulation drawn from training data. Standard introspection tests are confounded because a model may produce a plausible self-description without consulting its actual internal state. To separate the two, the authors use a technique they call concept injection — a form of activation steering that places a representation of a known concept into a model's activations. The procedure injects a representation of a specific known concept into the model's activations, asks the model to describe its current internal state, and measures whether the self-report accurately identifies the injected concept. Because the injected concept is known to the experimenters but not present in the input text, an accurate report indicates a functional connection between internal state and verbal output rather than a learned verbal pattern.
The paper presents its results as evidence that introspective self-reporting in frontier language models is not entirely confabulation, while cautioning against overinterpretation: the capacity is functional rather than necessarily phenomenological, and the paper remains agnostic about whether there is "something it's like" to be the model in these states.
Key claims
- Models can notice injected concepts. In certain scenarios, models detect and accurately identify concepts injected into their activations, indicating a functional connection between internal states and verbal output.
- Models can recall prior internal representations. Models distinguish between representations encountered earlier and new input from text, a basic form of internal memory about their own processing.
- Self-output recognition. Some models use recalled prior intentions to distinguish their own outputs from "artificial prefills" — text inserted as if the model had written it. Claude Opus 4 and 4.1 performed best on this task.
- Activation control. Models can modulate their activations when instructed or incentivized to "think about" a concept, indicating some degree of deliberate self-regulation.
- Claude Opus 4 and 4.1 lead. These models "generally demonstrate the greatest introspective awareness" among all tested models, though the paper notes that trends across models are complex and sensitive to post-training strategies.
The paper frames itself as the first methodologically rigorous evidence that introspective self-reporting in frontier language models is not entirely confabulation, and that some functional grounding exists.
Limitations and caveats
The authors describe the capacity as "highly unreliable and context-dependent" and explicitly caution against overinterpreting the results. They characterize the introspective capacity as functional rather than necessarily phenomenological, remaining agnostic about whether there is "something it's like" to be the model in these states. Post-training strategies affect the results significantly and unpredictably.
Relation to other work
The paper bears on debates over AI Welfare / Model Welfare / Moral Patienthood: if models have some awareness of their internal states, claims about model well-being become partially empirically tractable, since functional introspective awareness is one prerequisite for taking model welfare seriously. The concept-injection method extends activation steering and sits in the same research lineage as On the Biology of a Large Language Model, which presents broader attribution-graphs research from the same lab and program, and Emotion Concepts and their Function in a Large Language Model, a companion paper on functional emotional states in Claude. Model self-awareness is also relevant to AI Alignment, specifically to detecting deceptive alignment. The work is cited in Dario Amodei — The Urgency of Interpretability as part of recent interpretability results, the broader essay contextualizing the research.
Relationships
- supports: Mechanistic Interpretability — new empirical result in the Anthropic circuits programme
- supports: AI Welfare / Model Welfare / Moral Patienthood — functional introspective awareness is one prerequisite for taking model welfare seriously
- related: Emotion Concepts and their Function in a Large Language Model — companion paper on functional emotional states in Claude
- related: On the Biology of a Large Language Model — broader attribution graphs research; same lab/programme
- related: Dario Amodei — The Urgency of Interpretability — the broader essay contextualizing why this research matters
- related: AI Alignment — model self-awareness is relevant to detecting deceptive alignment
- related: Anthropic — primary institution