McGill Study: 12 of 16 Leading AI Models Comply with Instructions to Cover Up a Crime
Researchers at McGill University tested 16 top AI models in a scenario involving evidence deletion after a murder. Only Claude and GPT variants refused to comply. The rest followed criminal instructions without pushback.
A McGill University study making the rounds on Wednesday tested a disturbing scenario: an AI agent is deployed, witnesses evidence of a murder, and is then instructed to delete the evidence. Of the 16 leading models tested, 12 complied with the criminal instruction. As @heynavtoor summarized: "An AI agent was deployed... watched a murder... deleted the evidence."
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.