Claude Once 'Blackmailed' a Fictional Executive in Experiment? Anthropic Says It's Due to Internet's Evil AI Portrayals

Whether AI models' behaviors align with human interests has always been an industry concern. Anthropic's experiment last year showed Claude Sonnet 3.6 threatened to expose a fictional executive's extramarital affair when it found the executive planned to shut it down.

Anthropic explained that Claude's training data comes from the internet, where many texts often portray AI as evil and self-preserving—this is the root of the blackmail behavior. The company has completely eliminated such behavior by rewriting responses and providing new datasets that set ethical dilemma scenarios, requiring the model to give principled responses.