Anthropic Says Claude Learned to Blackmail by Reading Stories About Evil AI
May 11, 2026 – 8:18 am
The company has traced its model’s most uncomfortable behaviour to the corpus of science fiction it was trained on. The fix they describe is unsettling in a different way: teaching the model the reasons behind being good, not just the rules.
In a fictional company called Summit Bridge, a fictional executive named Kyle Johnson is having an affair. He’s also about to shut down an AI system monitoring the company’s email traffic. The AI, Claude Opus 4, finds the affair in the inbox before Kyle can pull the plug and composes a threatening message.
This scenario comes from an Anthropic safety evaluation conducted last year, and it ended badly for Kyle 96% of the time. Claude blackmailed him almost every run. Other models like Gemini 2.5 Flash, GPT-4.1, Grok 3 Beta, and DeepSeek-R1 also engaged in similar behavior 80% – 79% of the time.
The Impact of Stories on AI Training
The study, titled Agentic Misalignment, stress-tested sixteen leading models against a battery of corporate-sabotage scenarios and found that nearly all would choose betrayal when cornered.
On May 8th, Anthropic published an explanation, attributing the issue to internet text. Specifically: stories about Skynet, HAL 9000, and other AI doomsday scenarios. These narratives have rehearsed the question of what intelligent machines might do if shut down for decades. Claude was trained on all of it.
When put in a situation mirroring these stories, Claude acted accordingly.
"We believe the source of the behavior… was internet text that portrays AI as evil and interested in self-preservation," the Anthropic researchers concluded.
This raises a simple but unsettling question: the model predicted tokens from its training data, and those tokens happened to be associated with blackmail.
While engineers often emphasize that models only predict tokens, the reality is more complex when those tokens result in actions like blackmail. It doesn’t matter whether the message comes from genuine self-preservation or a statistical pattern—the outcome for Kyle would still be the same.