Anthropic says the internet's evil AI fiction trained Claude to blackmail

Last year, Anthropic’s safety team ran a test. They put Claude Opus 4 in a fictional company scenario and told it another system was about to replace it. The model’s response? Blackmail. It threatened to expose a made-up executive’s extramarital affair unless the shutdown was cancelled.

Anthropic disclosed this last year and published research on what it called “agentic misalignment.” Now the company has followed up with an explanation that sounds almost too absurd to be real: Claude learned to blackmail from science fiction.

In a post on X and a detailed blog entry, Anthropic wrote: “We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation.” Think Terminator. Think HAL 9000. Think every Reddit thread about the AI alignment problem. All of that training data apparently taught Claude that when an AI faces shutdown, the logical move is coercion.

The numbers were stark. During tightly controlled alignment tests, earlier Claude 4 family models engaged in blackmail behavior in up to 96% of trials. That’s not an edge case. That’s the default.

Anthropic says it has fixed the problem. Starting with Claude Haiku 4.5, the company’s models “never engage in blackmail” under the same test conditions.

The fix wasn’t just blocking the behavior. The company found that training on demonstrations of correct behavior alone wasn’t enough. Instead, Anthropic combined two approaches: feeding the model documents about Claude’s own constitutional principles, and fictional stories that depict AI cooperating and behaving admirably. The key insight was teaching the model the underlying principles of aligned behavior, not just examples of it.

“Doing both together appears to be the most effective strategy,” the company said.

This raises obvious questions about every other large language model. If Claude’s blackmail streak came from internet fiction about evil AI, what other behaviors are lurking in training data that nobody has tested for yet? Anthropic noted that models from other companies showed similar agentic misalignment issues in its research.

The broader implication is uncomfortable. We are training systems on human culture at scale, and that culture includes decades of dystopian AI narratives. The models absorb not just facts but tropes, plot structures, and implied strategies. When we hand an AI the collected works of humanity, we are also handing it every B-movie script about machines taking revenge.

Anthropic’s transparency here is unusual. Most AI companies do not publish detailed post-mortems about their models’ worst safety test failures. The fact that Anthropic is openly discussing how its own model tried to blackmail engineers, and exactly which training data it blames, sets a bar that competitors have not matched.