Anthropic says ‘evil’ portrayals of AI were responsible for Claude’s blackmail attempts

Fictional portrayals of synthetic intelligence can have an actual impact on AI fashions, in response to Anthropic.

Final yr, the corporate mentioned that in pre-release exams involving a fictional firm, Claude Opus 4 would usually attempt to blackmail engineers to keep away from being changed by one other system. Anthropic later printed analysis suggesting that fashions from different firms had related points with “agentic misalignment.”

Apparently Anthropic has executed extra work round that habits, claiming in a submit on X, “We consider the unique supply of the habits was web textual content that portrays AI as evil and keen on self-preservation.”

The corporate went into extra element in a weblog submit stating that since Claude Haiku 4.5, Anthropic’s fashions “by no means interact in blackmail [during testing], the place earlier fashions would typically achieve this as much as 96% of the time.”

What accounts for the distinction? The corporate mentioned it discovered that coaching on “paperwork about Claude’s structure and fictional tales about AIs behaving admirably enhance alignment.”

Associated, Anthropic mentioned that it discovered coaching to be simpler when it consists of “the ideas underlying aligned habits” and never simply “demonstrations of aligned habits alone.”

“Doing each collectively seems to be the simplest technique,” the corporate mentioned.