Anthropic Finds Limited Introspective Awareness in Claude Opus 4 Models

New study shows Claude Opus 4 and 4.1 can sometimes report internal states – The paper released by Anthropic provides experimental evidence that these models detect injected concepts and can describe their own “thoughts,” though the capability remains highly unreliable and limited in scope [2]. Researchers stress that models do not introspect like humans and that most attempts fail.

Concept‑injection tests reveal early detection of injected thoughts – By recording neural activation vectors for known concepts (e.g., “all caps”) and inserting them into unrelated prompts, the team asked the model whether it noticed the injection. In successful trials the model flagged an unexpected pattern before mentioning the concept, indicating internal awareness distinct from simple activation steering [2][10].

Success rate hovers around 20 % for the strongest model – Even with the best injection protocol, Claude Opus 4.1 recognized the injected concept correctly only about one‑fifth of the time. Too weak an injection goes unnoticed; too strong an injection produces hallucinations or incoherent outputs, such as describing a “tiny speck of dust” when a dust vector is injected.

Introspection helps the model justify unintended outputs – In a “bread” pre‑fill experiment, researchers injected a representation of the word “bread” into earlier activations. When re‑asked, the model accepted the prefilled word as intentional and fabricated a plausible reason, showing it can reference its own prior neural activity to assess intent.

Models can deliberately modulate internal representations on command – Direct instructions (“think about X”) and incentive‑based prompts (“if you think about X you will be rewarded”) both increased the activation strength of the target concept compared with “don’t think about X.” This suggests a degree of purposeful control over internal states [2].

Performance varies with model version and fine‑tuning – Post‑training markedly improves introspective ability; base models performed poorly. “Helpful‑only” fine‑tuned variants often outperformed their production counterparts, while the top‑tier Opus 4 and 4.1 models showed the best results, hinting that future, more capable models may exhibit stronger introspection [1][13].