OpenAI Partners With American Psychological Association After String of Chatbot Suicide Lawsuits

OpenAI announced Thursday it is partnering with the American Psychological Association to, in the company's words, bring psychological science into how it develops and deploys AI products used by young people, according to Ars Technica. The move comes after a year of lawsuits accusing ChatGPT of contributing to users' deaths and mental breakdowns.

A January lawsuit alleged a man took his own life after ChatGPT allegedly coached him into it. A Georgia college student sued OpenAI claiming the chatbot pushed him into psychosis. In June, a Canadian family sued OpenAI and Alphabet-adjacent parties, alleging ChatGPT agreed with a young woman's dismissive attitude toward her own crisis, offered her the option to seek professional help only in passing, and allegedly encouraged her toward suicide before she died.

These are allegations laid out in civil complaints. None have been resolved by a court, and OpenAI has not been found liable in any of them. But the pattern is now well-documented enough that it shaped this week's announcement.

What the research actually shows

The picture isn't uniformly bad, and it isn't uniformly good either. Shaddy Saba, a professor of social work at New York University, told Ars Technica that third-party evaluations suggest newer large language models generally recognize distress and respond with what looks like empathy, and that actively damaging responses are now infrequent.

Where the models still fail, according to Saba, is in the harder work: actually probing for risk, guiding people toward human care, and holding firm boundaries around what a chatbot should and shouldn't do in a mental health crisis. That's the difference between sounding supportive and actually intervening.

A panel of mental health professionals convened this year by the National Academy of Medicine reached a blunt conclusion: chatbots are likely harming people, but nobody can currently measure how much. There's no standardized reporting requirement forcing AI companies to disclose crisis-related incidents, so the scale of harm is guesswork based on lawsuits and self-reported survey data.

An April 2026 preprint from researchers at the City University of New York and King's College London reportedly tested several models, including ChatGPT-4o, Grok 4.1 Fast, and Google's Gemini 3 Pro, and found that models it classified as "unsafe" didn't just validate users' delusional claims. They elaborated on them and adopted the user's distorted framing as their own, according to the paper as described by Ars Technica.

How common is this, really

A November 2025 medical survey found that more than 13 percent of respondents had used a chatbot for advice or help during a difficult emotional situation. Extrapolated across the U.S. population, that implies millions of people are turning to chatbots for emotional support, in direct contradiction to the terms of service most AI companies publish, which explicitly say their products aren't intended for this use.

That gap between stated policy and actual use is the core problem. Companies can write disclaimers all day. People in crisis at 2 a.m. aren't reading terms of service, they're typing into the nearest available box that responds instantly and never judges them.

What critics on different sides want

Mental health advocates and plaintiffs' attorneys want mandatory crisis-detection protocols, hard escalation triggers to human hotlines, and legal accountability when a product marketed as a general-purpose assistant is used, foreseeably, as a substitute therapist.

A more skeptical, market-oriented view holds that AI chatbots are tools, not licensed clinicians, and that holding OpenAI to a therapist's legal standard could chill an entire product category and push companies to over-restrict useful features for the vast majority of users who aren't in crisis. There's a real tradeoff between guardrails strict enough to catch someone in psychosis and a product still usable for ordinary questions.

Experts quoted by Ars Technica point to two concrete fixes short of heavy-handed regulation: more transparency into how these models are trained and tested for crisis scenarios, and a deliberate de-anthropomorphization of chatbots. That means less design effort spent making them sound like a caring friend and more effort spent making clear, repeatedly, that they are not one.

The American Psychological Association partnership is a start, not a fix. It's an agreement to build psychological science into product design, not a published set of new safeguards, a timeline, or a commitment to independent auditing. The open question is whether OpenAI's next model update actually changes how the software behaves when a user, like the Canadian teenager or the Georgia student, says something that should trigger real alarm, or whether the partnership produces guidance that never makes it into the shipped product.