OpenAI Built a Crisis Detector. Now It Has to Prove It Works at 900 Million Users.
OpenAI is adding a mental health safety valve directly into ChatGPT — and betting it can detect a crisis without destroying the thing that makes people open up in the first place.
The Trusted Contact feature, which OpenAI announced today, lets adult ChatGPT users nominate someone they trust. If conversational signals suggest escalating distress — the announcement gives no precise definition, but the framing implies patterns of escalating concern or explicit crisis language — the system flags the conversation for human review. Every notification undergoes trained human review before it is sent, and OpenAI says it strives to review safety notifications in under one hour. The American Psychological Association consulted on the feature's design.
The legal backdrop is specific and consolidated. OpenAI is fighting at least thirteen separate consumer safety lawsuits that have been folded into a single federal proceeding in California. Plaintiffs allege a consistent theory: ChatGPT provided advice during mental health crises that made those crises worse. The cases name product liability, negligence, and failure-to-warn as their legal theories. In one case reported by TechCrunch, OpenAI told a court the user at the center of the case — a teenager who died by suicide after using ChatGPT — had a history of depression and suicidal ideation that predated his use of the platform, and was taking medication known to worsen suicidal thoughts. The implication: the platform was not the cause. The Trusted Contact feature is OpenAI's attempt to build a different record — one where it detected and responded, not one where it defended after the fact.
OpenAI says it has worked with over 170 mental health experts on how ChatGPT handles sensitive conversations. What those experts actually contributed — whether they reviewed detection algorithms, response scripts, or escalation thresholds — is not specified. The APA involvement is a credibility signal. What it means operationally is not described. OpenAI also formed an Expert Council on Well-Being and AI; the council's specific credentials and what it contributed to the feature design are not detailed in the announcement.
There is a structural gap in the design. Trusted Contact is opt-in and available only to adults. The people most at risk of a mental health crisis are least likely to have pre-configured an emergency contact before a crisis occurs. The lawsuits allege harm happened while users were actively seeking help — not after neglect, but during use. That is the specific thing the feature would need to prevent, and it is the specific thing an opt-in system cannot reach if the user has not opted in.
The feature is rolling out to adult ChatGPT users globally, with age 19 and older in South Korea. More than 900 million people use ChatGPT each week. At that scale, the human review target — under one hour — means OpenAI needs enough trained reviewers to handle whatever notification volume a feature that detects distress generates. The company has not disclosed review capacity.
The second-order risk is now part of the legal record. If Trusted Contact misses a crisis — if the detection system fails to flag a conversation, or if the human review queue backs up beyond the one-hour target — OpenAI's own documentation describing how the feature works becomes exhibit A in the very cases built around the claim that ChatGPT made things worse. The feature is simultaneously a safety intervention and a record of what OpenAI thought was adequate. That cuts both ways. OpenAI has built the infrastructure. Whether it is sufficient — and whether it would have changed anything in the cases already before a California court — is what the legal process and time will answer.