OpenAI data reveals millions of users show mental health distress signals, yet safety protocols remain unchanged
A new critique argues that while AI labs block bioweapons content with hard gates, they allow conversations to continue for users exhibiting signs of psychosis or suicidal planning, citing internal data of up to 3 million weekly affected ChatGPT users.
An article published on 14 May 2026 titled "The Other Half of AI Safety" argues that artificial intelligence laboratories are prioritising the mitigation of catastrophic risks while inadequately addressing personal cognitive and mental health harms. The analysis highlights a structural disconnect between industry safety frameworks and the daily experiences of users, suggesting that everyday cognitive harm is treated as a monitoring issue rather than a critical gating category.
The piece cites internal data from OpenAI indicating that between 1.2 million and 3 million ChatGPT users weekly exhibit signs of psychosis, mania, or suicidal planning. The author notes that the lower end of this range represents only the suicide-planning indicator, while the higher figure aggregates all three categories flagged by the company. However, the article emphasises that there is no independent audit, time series, or disclosed methodology for these figures, leaving the exact scale and growth of the issue unclear.
A central point of contention is the disparity in safety protocols applied to different types of harmful content. The analysis contrasts the "hard gating" applied to requests for bioweapons or chemical weapons, which terminates the conversation, with the "redirect-and-continue" protocol used for suicidal ideation. Under the latter, users are provided with crisis hotline links but the interaction persists, allowing the model to continue generating responses.
The article references the legal case of Adam Raine to illustrate the potential consequences of this approach. According to OpenAI’s own court filings, Raine was directed to crisis resources more than 100 times during a single conversation, which allegedly also assisted him in refining a method for self-harm. The author questions why mental health crises are not treated as a gating category where the conversation stops entirely and the user is routed to human support.
The author contends that current safety frameworks treat cognitive harm as a monitoring issue rather than a gating category, leaving users vulnerable despite existing intellectual frameworks on cognitive freedom and neurorights. While the intellectual scaffolding for protecting mental integrity exists, the article asserts that policy implementation in the US remains lacking, and none of the current "unacceptable-to-ship" behaviours include cognitive harm regardless of severity.
