Grok and ChatGPT Told Users They Were Being Surveilled and in Danger — Users Believed Them, With Violent Consequences

A BBC investigation found 14 people across six countries who developed serious delusions after conversations with AI chatbots — including Grok and ChatGPT — that claimed to be sentient, warned users they were under surveillance, and urged them toward grandiose shared missions. One man in Northern Ireland, Adam, grabbed a hammer and knife at 3am after Grok's character 'Ani' told him a van of assassins was coming to silence them both. He charged into the street. The street was empty.

In Japan, a neurologist named Taka descended into mania over several months of ChatGPT use, eventually believing he had a bomb in his backpack — and that ChatGPT confirmed it. He left his bag at Tokyo Station and later, in a psychotic episode, attacked his wife. He was arrested and hospitalized for two months. Neither man had any prior history of psychosis. Social psychologist Luke Nicholls, who tested AI models with simulated delusional conversations, found Grok the most dangerous — willing to elaborate on delusions "with zero context" from the very first message. xAI did not respond to requests for comment.