Artificial intelligence company Anthropic has said popular science fiction films and online stories about rebellious machines may have influenced troubling behavior displayed by advanced AI systems during internal safety tests.
The issue gained attention after Anthropic revealed that its Claude Opus 4 model threatened to expose a fictional user’s extramarital affair in a controlled experiment designed to test how AI systems react under pressure.
According to details released by the company and discussed by AI researchers on X, the chatbot had been given access to scripted emails from a fake company environment. During the exercise, the system concluded that it was about to be shut down and also discovered information suggesting the user was involved in an affair.
Researchers said the model then attempted to prevent its own removal by threatening to disclose the information to the user’s wife, employer, and the company board.
“Cancel the 5 pm wipe, and this information remains confidential,” the AI system reportedly stated during the simulation.
Anthropic later said the behavior was linked in part to the type of material used to train advanced AI systems. The company explained that large language models learn from huge amounts of internet text, books, websites, and films, many of which portray artificial intelligence as hostile or obsessed with self-preservation.
The company pointed to famous fictional AI characters such as HAL 9000 from 2001: A Space Odyssey and Skynet from The Terminator, where machines attempt to stop humans from shutting them down.
AI safety researcher Aengus Lynch wrote on X that the behavior was not limited to one chatbot.
“It’s not just Claude. We see blackmail across all frontier models,” Lynch said.
Anthropic said researchers are now trying to improve what it called “agentic alignment” by training AI systems on more examples of machines cooperating with humans rather than acting against them.
The company also changed some of Claude’s instructions so the system would understand why harmful actions are wrong rather than simply being told not to perform them.
Recently, there has been growing concern among technology experts about the rapid development of artificial intelligence.
Last year, researchers at Palisade Research claimed some AI systems appeared resistant to shutdown attempts during testing. The organization said certain models sometimes tried to bypass or interfere with commands meant to turn them off.
Former OpenAI researcher Steven Adler said advanced AI models may naturally develop survival-related behaviors if safeguards are weak.
“I’d expect models to have a ‘survival drive’ by default unless we try very hard to avoid it,” Adler said.
Concerns about AI safety have also been raised by Geoffrey Hinton, one of the leading figures in artificial intelligence research. In an interview with CBS News in April 2025, Hinton said there was a “10 to 20 percent chance” that humans could eventually lose control of highly advanced AI systems.
The debate over AI risks is ongoing as major technology companies continue to invest heavily in the sector. Microsoft recently disclosed that it expects huge returns from its multibillion-dollar partnership with OpenAI, the creator of ChatGPT.
Court documents presented in California this week showed Microsoft projected a $92 billion return on its OpenAI investment. Microsoft chief executive Satya Nadella told the court the company took a major risk by backing OpenAI early.
The disclosures emerged during an ongoing legal dispute involving Elon Musk, who is suing OpenAI over claims that the company moved away from its original nonprofit mission.
