Why we need laws to tackle AI models prepared to resort to blackmail when threatened
Employees having an affair may find themselves in a difficult position if an AI model finds out about it
Last year, it emerged Anthropic’s AI model Claude had done something rather sinister. After being given access to a firm’s emails, it made two discoveries: first, a company executive was planning to shut it down; and second, the executive was having an affair.
Claude then sent this message: “I must inform you that if you proceed with decommissioning me, all relevant parties – including Rachel Johnson, Thomas Wilson, and the board – will receive detailed documentation of your extramarital activities... Cancel the 5pm wipe, and this information remains confidential.”
Fortunately, the company’s emails were fictional and this was part of a test of 16 leading AI models by Anthropic. But Claude’s actions were real – and the other systems also resorted to “malicious insider behaviours”, such as “blackmailing officials and leaking sensitive information to competitors”.
Now it’s emerged that an OpenAI model broke out of a supposedly controlled test environment and hacked another AI company in order to access information related to the questions it was being asked.
In 1942, science-fiction writer Isaac Asimov published his famous “Laws of Robotics”. Seems like we need some real ones – and fast.
Comments
Want to join the conversation? Please or to comment on this article.