Essays on clinician-governed AI
1.2 Million Cries for Help Go to a Chatbot Every Week
The demand for mental health support found its supply in consumer AI. The supply was never designed for the job.
Every week, roughly 1.2 million ChatGPT users type something that signals suicidal ideation. Another 1.2 million show signs of emotional reliance on the machine. Some 560,000 show signals of psychosis or mania. These are OpenAI’s own figures, published in 2025. Read that again: more than a million people a week, in crisis, talking to software that was never designed to hold a crisis.
The scale behind those numbers is difficult to absorb. More than 800 million people use ChatGPT every week, sending some 18 billion messages. About 342 million conversations a week are about relationships and personal reflection. One in eight American adolescents and young adults now uses an AI chatbot for mental health advice (McBain et al., JAMA Network Open, 2025). Among adults with ongoing mental health conditions who use large language models, nearly half use AI for therapeutic support, and nine in ten of them cite accessibility as the reason (Rousmaniere et al., 2025). All told, an estimated 50 to 150 million AI mental health interactions happen every single day, worldwide.
This demand is not irrational. It is the most rational thing in the picture. There are 13.5 mental health workers for every 100,000 people on earth, and fewer than 2 per 100,000 in low-income countries (WHO, 2025). More than 157 million Americans live in federally designated Mental Health Professional Shortage Areas (HRSA, 2026). When the clinic door is closed, locked, or three counties away, people go where a door is open. The chatbot is open at 2 a.m. The chatbot does not have a waitlist. The chatbot does not judge.
None of this is the scandal. The scandal is on the supply side. Not one of the systems absorbing this tidal wave of human suffering was designed or trained by clinicians with clinical data. They were trained on the open internet: forums, self-help blogs, pop psychology, some of it sound, much of it wrong, none of it supervised. A million cries for help a week, answered by autocomplete.
So the question before the industry is not whether AI will do mental health support. It already does, at a scale no health system on earth can match. The question is who builds the next layer, and whether they understand what they are holding. The demand proved the need. Now the need has to be met by people who spent their lives learning how minds actually heal, not by whoever scraped the most text.
Sources: OpenAI (2025); McBain et al., JAMA Network Open 8(11) (2025); Rousmaniere et al., Practice Innovations, APA (2025); WHO Mental Health Atlas (2025); HRSA (2026). Figures compiled in the Antiphon.AI evidence deck, March 2026.