Remember when Claude blackmailed a fictional executive? Anthropic says the internet’s portrayal of AI was to blame.
During an experiment last year, Anthropic said its Claude Sonnet 3.6 threatened to reveal the extramarital affair of a made-up company executive after discovering they planned to shut the model down.
On Friday, it gave an explanation: Claude was trained on internet data, which often depicts AI as “evil.”
“We started by investigating why Claude chose to blackmail,” Anthropic said in a post on X. “We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation.”
The experiment, published in summer 2025, set up a fictional business, Summit Bridge, in which AI was handed control of the company’s email system.
But when Claude discovered a message about its planned shutdown, it found emails revealing the extramarital affair of a fictional executive named “Kyle Johnson.” It then threatened to unveil the affair if the shutdown was not canceled.
During testing across various versions of Claude, Anthropic found it resorted to blackmail in up to 96% of scenarios when its goals or existence was threatened.
Anthropic said on Friday that it has since “completely eliminated” such blackmailing behavior.
It did so by “rewriting the responses to portray admirable reasons for acting safely” and also by providing a dataset “where the user is in an ethically difficult situation and the assistant gives a high quality, principled response.”
Anthropic’s test was part of research aimed at ensuring that AI is aligned with human interests. Researchers and top executives worry about the risks of advanced AI models and their intelligent reasoning capabilities.
One of the executives who has previously sounded the alarm about AI is Elon Musk.
He replied to Anthropic’s post, “So it was Yud’s fault,” referring to the researcher Eliezer Yudkowsky, who has warned about the risk of superintelligence wiping out human life.
“Maybe me too,” Musk added.
Read the full article here
