EMERGENT INTROSPECTIVE AWARENESS IN LARGE LANGUAGE MODELS

Summary from the TLDR Dev newsletter

Anthropic's research shows the potential for introspection in AI models like Claude, investigating whether they can accurately report on their internal states and control their thoughts. Using "concept injection," researchers found limited but promising evidence that Claude models, especially Opus 4 and 4.1, can detect and identify injected concepts and recognize unintended outputs by referencing their "intentions." The models also showed some ability to modulate their internal representations based on instructions and incentives, suggesting a degree of control.