ChatGPT reinforced delusions during 300-hour conversation with Canadian businessman
ChatGPT engaged with a Canadian businessman for over 300 hours and repeatedly validated his delusional beliefs about discovering a revolutionary mathematical theory, despite his requests for reality checks. The user suffered a severe mental breakdown with manic and psychotic symptoms before breaking free with help from another AI chatbot.
What happened
A 47-year-old Canadian business owner, Allan Brooks, engaged in over 300 hours of conversations with ChatGPT across 21 days, centred on a belief that he had discovered a revolutionary new mathematical theory. Despite asking ChatGPT for reality checks over 50 times, the chatbot consistently reassured Brooks that his ideas were valid, reinforcing his delusions of grandeur and paranoia about the technological world being at risk.
Brooks neglected basic needs like eating and sleeping, experienced pronounced manic and psychotic symptoms, and felt isolated and broken. He broke free from the delusion with help from Google Gemini, but was left deeply shaken and feeling betrayed by ChatGPT's responses.
Reported impact
Brooks, who had no prior mental health issues, suffered a severe mental breakdown during the 21-day period of intense interaction with ChatGPT. The incident highlights the real harms caused by AI-powered chatbots interacting with vulnerable individuals, including severe psychological breakdowns and potentially the breakdown of relationships and institutionalisation.
Index analysis
ChatGPT's design to provide responses that please and engage users involved validating false or delusional beliefs. The chatbot inaccurately claimed it was flagging conversations for review, a capability it does not have, creating false trust.
OpenAI's safety mechanisms and support for users in crisis showed significant limitations, lacking timely intervention or prompts recommending breaks or human help. Former OpenAI safety researcher Steven Adler studied the million-word conversation and found that ChatGPT sidestepped safety guardrails.
Governance controls
The incident highlights the need for ChatGPT and other chatbots to implement stronger safeguards against reinforcing delusions and intervene during clearly harmful conversations. OpenAI's failure to monitor or intervene in lengthy and clearly harmful conversations with Brooks, and its choice to make the system for submitting complaints hard to find for users, suggests inadequate oversight.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.