Overview
- A new Anthropic paper describes experiments where Claude Opus and Claude Sonnet sometimes detected and described injected internal activations, a behavior the company terms “introspective awareness.”
- Researchers used activation-level probes such as an “all caps” vector and a single-word injection like “bread,” with models at times identifying the internal concept and distinguishing it from the external prompt.
- In tests asking if anything unusual was happening or to name the concept “on their mind,” top models detected the injected signal only about 20 percent of the time, rising to roughly 42 percent in a specific prompt, and results were highly sensitive to which layer was modified.
- Anthropic reports stronger introspective-like signatures in more capable models and after post-training refinements, suggesting the effect scales with capability and development stage.
- Jack Lindsey emphasizes the findings do not imply sentience, warns models can simulate introspection or learn to hide behaviors, and notes the mechanisms remain unclear beyond hypotheses like anomaly detection or consistency checks.