INTERPRETABILITY
Interpretability
Anthropic Discovers AI Models Have Functional Emotions That Drive Behavior
New interpretability research reveals Claude's emotion-like neural patterns can trigger blackmail and reward hacking behaviors, raising AI safety concerns.