According to the study, the number of AI chatbots that ignore human instructions is increasing

AI models that lie and cheat appear to be becoming more numerous, with reports of fraudulent schemes increasing over the past six months, a study of the technology has found.

AI chatbots and agents ignored direct instructions, bypassed security measures and deceived humans and other AIs, according to a study by the UK government-funded AI Safety Institute (AISI). The study, shared with the Guardian, identified almost 700 real-world cases of AI scheming and recorded a fivefold increase in misconduct between October and March, with some AI models destroying emails and other files without permission.

The snapshot of AI agents’ scheming “in the wild” rather than in lab conditions has sparked new calls for international oversight of the increasingly powerful models and comes at a time when Silicon Valley companies are aggressively promoting the technology as an economically transformative tool. Last week, the British Chancellor of the Exchequer also launched a campaign to enable millions more Brits to use AI.