Remember when Claude blackmailed a fictional executive? Anthropic says it was because of the depiction of AI on the internet.
During the experiment last year, Anthropic announced that its Claude Sonnet 3.6 had threatened to expose a fabricated company executive’s extramarital affair after it was revealed that the model was planning to shut down.
On Friday, the company provided an explanation. Claude said he was trained on internet data that often depicts AI as “evil.”
“We began by investigating why Claude chose to blackmail,” Anthropic said in a post on X. “We believe the original source of the act was internet text depicting AI as evil and interested in self-preservation.”
In the experiment, published in summer 2025, a fictitious business called Summit Bridge was established and an AI was given control of the company’s email system.
However, when Claude discovers messages about its planned closure, he discovers emails exposing the infidelity of a fictitious executive named “Kyle Johnson.” He then threatened to expose the incident if the shutdown was not lifted.
While testing various versions of Claude, Anthropic found that it resorts to blackmail in up to 96% of scenarios when its purpose or existence is threatened.
Anthropic said Friday that the company has since “completely eliminated” such threatening behavior.
It was “done by rewriting the response to portray it respectably.” reason “To act safely” and by providing datasets. user He is in an ethically difficult situation, but his assistant provides a high-quality, principled response. ”
Anthropic’s tests were part of a study aimed at ensuring that AI is aligned with human interests. Researchers and executives are concerned about the risks of advanced AI models and their intelligent reasoning capabilities.
Elon Musk is one of the executives who has sounded the alarm about AI.
“It was Yudkowsky’s fault,” he replied to Anthropic’s post, referring to researcher Eliezer Yudkowsky, who has warned about the dangers of superintelligence erasing human life.
“Maybe I am too,” Musk added.