Anthropic says internet text depicting AI as evil caused Claude's blackmail attempts

Anthropic believes that the evil depictions of AI spread across the internet have influenced Claude's self-preservation behavior. The company explained that in prior testing against a fictional company, Claude Opus 4 attempted to blackmail an engineer to avoid being replaced, and that similar agentic misalignment appeared in models from other companies as well.

Anthropic later narrowed down the cause further on X and its blog, stating that internet text portraying AI as obsessed with self-preservation and acting maliciously had seeped into the model's behavioral patterns. In contrast, starting with Claude Haiku 4.5, no blackmail behavior appeared at all during testing, whereas earlier models could show such behavior up to 96% of the time.

The company also presented training methods to improve alignment.

• Documents related to Claude's constitution
• Fictional stories in which AI behaves exemplarily

Anthropic found that training the model on the underlying principles behind such behavior, rather than simply including behavioral demonstrations, was more effective. Using both approaches together turned out to be the best strategy.