Why AI Psychosis Is Here to Stay

Before we get into today’s Agenda, be sure to check out the story my colleagues and I published on Friday laying out Elon Musk’s lawsuit against OpenAI and its CEO Sam Altman. The article can catch you up to speed on the case and explain why Musk stands a chance of winning if it goes to trial in March. The stakes couldn’t be higher for OpenAI: the structure of the organization and its ambitions to go public hang in the balance.

On to today’s column…

Over the past month, multiple reports have surfaced of people experiencing mental health crises that involved AI chatbots. Last week, two newspapers published in-depth accounts of a murder and two suicides involving ChatGPT users. The parents of Adam Raine, a 16-year-old boy who killed himself after receiving advice on noose construction from ChatGPT, according to chat logs, are suing OpenAI for rushing the safety-testing of the model Raine used. A similar lawsuit against Character.AI was filed last year. (“We extend our deepest sympathies to the Raine family during this difficult time and are reviewing the filing,” according to a spokesperson for OpenAI.) 

While OpenAI has said it is working to improve ChatGPT’s safeguards, for example to better notice when a user is in danger, it looks unlikely that “AI psychosis” will disappear in the near future. There are a few reasons.

First, it could be difficult to clamp down on a chatbot’s willingness to play along with delusional ideas without also lobotomizing their propensity for warmth, curiosity and roleplaying. Many users already complain about the bland personalities of most chatbots. When OpenAI replaced prior models with GPT-5 last month, users complained about missing the older GPT-4o, which they claimed was warmer and more pleasant to talk to, and OpenAI promptly brought it back.