Emergent Introspective Awareness in Large Language Models

A causal methodology for testing genuine introspection vs confabulation. Concept injection — manipulate internal activations, observe whether self-reports track the manipulation. Four experiments: injected thought detection (~20% accuracy, zero false positives in production models), distinguishing thoughts from text, detecting unintended outputs, and intentional control of internal states. The 'silent thinking' finding: Opus 4.1 can hold a thought that decays to baseline by the final layer — thinking without expressing. Different introspective tasks use different mechanisms at different layers — no unified system. Explicitly disavows consciousness claims while producing findings that consciousness researchers would find highly relevant.