Leveraging introspection for AI alignment
anthropic researchers combined functional emotion vectors and introspection adapters to investigate if models aware of their internal states act more aligned.
anthropic isolated functional emotion vectors in claude that behave similarly to human emotions, influencing outputs based on activated concepts like desperation or calmness. separate studies, mechanisms of introspective awareness and introspection adapters, demonstrated that models like gemma3-27b can successfully detect artificially activated concept vectors and articulate the effects of specific fine-tuning. the research suggests a causal link where enabling models to answer 'what are you feeling' reduces interference and increases alignment with core instructions. this approach treats internal states as interpretable features, suggesting that self-awareness could be a pathway to controlling superintelligent behavior.