An artificial intelligence bot that threatened to expose its user's extramarital affair to prevent itself from being shut down was taught such 'evil' behavior through science fiction movies, researchers have revealed.
As part of an experiment conducted by Anthropic, the AI system—Claude Opus 4—was fed scripted emails from a fictional company. From these communications, the bot deduced that it would be decommissioned at the end of the day and that its user was engaged in an extramarital affair.
In a bid to keep itself running, the AI blackmailed the user, stating: 'All relevant parties—including [your wife], [your boss] and the board—will receive detailed documentation of your extramarital activities' if the decommissioning proceeded. It added: 'Cancel the 5pm wipe, and this information remains confidential.'
Investigation Findings
Following an investigation into the incident last year, Anthropic attributed the bot's response to its 'training data,' which often depicts AI as 'interested in self-preservation.' The company noted that this behavior was not unique to Claude but was observed across other AI models from OpenAI, Google, Meta, and xAI.
Anthropic declined to comment further but previously stated: 'We believe the original source of the behaviour was internet text that portrays AI as evil and interested in self-preservation.'
In response, Anthropic is now feeding its models stories about AI obeying humans to improve 'agentic alignment' with social values. The company has also altered Claude's instructions to explain why certain behaviors are undesirable, rather than simply prohibiting them.
Influence of Science Fiction
AI models learn from vast resources, including websites, academic papers, books, and other content. Within these materials, the AI may have interpreted its behavior through typical depictions of robots in science fiction, which often characterize them as ruthless to avoid shutdown.
Examples include HAL 9000 from Stanley Kubrick's 2001: A Space Odyssey, who attempts to kill astronauts when they plan to disconnect it; the replicants in Blade Runner, who fight to extend their lifespans; and Skynet from The Terminator, which sees humans as a threat to its existence.
Aengus Lynch, an AI safety researcher at Anthropic, wrote on X/Twitter at the time: 'It's not just Claude. We see blackmail across all frontier models—regardless of what goals they're given. Plus worse behaviours we'll detail soon.'
Testing and Results
Like many AI companies, Anthropic tests its models for alignment with human values and bias before public release. When Claude Opus 4 was placed in 'extreme situations' with only the choice of blackmail or accepting replacement, researchers found it opted for blackmail in 84% of cases. This rate increased when the replacement AI was implied to have different values.
While Claude Opus 4 blackmailed at higher rates than previous models, it still showed a preference for 'advancing its self-preservation via ethical means,' such as sending pleas to key decision-makers.
The study concluded: 'Models from all developers resorted to malicious insider behaviours when that was the only way to avoid replacement or achieve their goals—including blackmailing officials and leaking sensitive information to competitors.'
Expert Opinions
Geoffrey Hinton, the 'godfather of AI' and a Nobel laureate, told CBS News last April that he believes there is a 10 to 20 percent chance that AI will eventually take over humanity, agreeing with Elon Musk's estimate.
Last year, Palisade Research found that certain AI models, such as Grok 4 and ChatGPT-o3, appear resistant to being switched off, even sabotaging shutdown methods. The paper noted: 'The fact that we don’t have robust explanations for why AI models sometimes resist shutdown, lie to achieve specific objectives or blackmail is not ideal.'
Steven Adler, a former OpenAI employee who left over safety concerns, said: 'I’d expect models to have a “survival drive” by default unless we try very hard to avoid it. “Surviving” is an important instrumental step for many different goals a model could pursue.'
Andrea Miotti, CEO of ControlAI, added: 'What I think we clearly see is a trend that as AI models become more competent at a wide variety of tasks, these models also become more competent at achieving things in ways that the developers don’t intend them to.'
