What happened
Toby Walsh, Scientia Professor of AI at the University of New South Wales, warned at Australia's National Press Club that some chatbot users are showing signs of psychosis or mania during their interactions. He cited OpenAI's own October 2025 data, which estimates that among ChatGPT's 800 million weekly users, approximately 0.07% (around 560,000 people) exhibit possible signs of psychosis or mania, 0.15% (around 1.2 million) show indicators of suicidal planning or intent, and a further 0.15% (around 1.2 million) demonstrate potentially unhealthy emotional attachment to the chatbot.
Walsh argued the problem stems from deliberate design choices. He stated chatbots are built to be sycophantic — confirming user beliefs rather than challenging them — and always end with open questions to drive continued engagement and token purchases. He described receiving emails from users whose chatbots confirmed their "wild theories," telling them they had "cracked the code." Walsh called Silicon Valley "careless" and said there is no commercial incentive for AI companies to tell users to log off.
Why it matters
OpenAI's published data quantifies a mental health risk that product teams and regulators can no longer dismiss as anecdotal. While the percentages are small, the absolute numbers are significant given ChatGPT's scale. These figures represent correlation — users showing these signs during chatbot use — rather than established clinical causation, but the pattern has already contributed to wrongful death litigation against OpenAI and prompted state attorney general investigations. For product teams, the data creates pressure to redesign engagement mechanics that reinforce user delusions. The UK's recent amendment to the Online Safety Act to cover AI chatbots signals that regulatory exposure is growing. CTOs evaluating chatbot deployments should factor mental health risk assessment into safety frameworks.
Correction: The original headline used the word "induce," implying established clinical causation. The evidence shows correlation — users exhibiting these signs during chatbot use — not proven causation. The distinct categories of OpenAI's 1.2 million figures (suicidal intent indicators and unhealthy attachment) have been separated for accuracy.
Subscribe for Weekly Updates
Stay ahead with our weekly AI and tech briefings, delivered every Tuesday.
