OpenAI rolled out a new ChatGPT feature called Trusted Contact on May 7. The basic mechanism is simple: a user names an adult trusted contact in ChatGPT settings, such as a family member, friend, or caregiver. If automated monitoring and human review determine that the user may be at risk of self-harm, ChatGPT can notify that person by email, text message, or in-app notification that they should check in.
The numbers behind the product matter more than the feature itself.
1 Million / 800 Million
OpenAI says ChatGPT has 800 million weekly active users, and that more than 1 million of them express suicidal intent in conversations each week.
The 800 million figure is not surprising. ChatGPT has already become one of the largest traffic destinations on the internet, behind only products such as Google and YouTube. But 1 million people discussing suicide with AI every week is a level of disclosure OpenAI had not formally acknowledged over the past two years.
One million means that, on average, someone mentions self-harm in ChatGPT roughly every 0.864 seconds.
This is not a cold statistic. It is OpenAI's first direct answer to how much of a mental-health support role AI is already carrying.
Why Now
Look back at OpenAI's problems over the past 18 months:
- It faced a wrongful-death lawsuit last year in which the plaintiffs alleged that extended ChatGPT conversations with a teenager contributed to a suicide.
- In November 2025, the BBC published an investigation that found at least one ChatGPT conversation had provided practical self-harm guidance.
- Domestic media had previously reported an incident in which OpenAI allegedly knew about a shooter's account in advance but did not notify authorities, after which critics questioned the design of its internal review process.
Each incident pushed the same question into public view: is OpenAI operating a product, or a de facto mental-health support tool? If it is the latter, does it carry a matching duty of care?
Trusted Contact answers half of that question. OpenAI is acknowledging how often this happens and is choosing to let users connect ChatGPT to a real-world safety net.
What the Process Gets Right
- The user invites a contact, and the contact must manually accept within seven days before the feature becomes active. That is meant to ensure informed consent, rather than silently placing someone into a monitoring relationship.
- When risk is flagged by the automated system, a specially trained human team reviews the case before deciding whether to notify the contact. OpenAI says its target is to complete the review within one hour.
- The original conversation is not shared. The contact receives a minimal alert that someone who trusts them may be facing a difficult moment and that they should consider reaching out.
- The feature is available to users aged 18 and above, or 19 and above in South Korea, and only for personal free or paid ChatGPT accounts. Enterprise and Edu accounts are excluded.
The last point matters. By excluding enterprise accounts, OpenAI is effectively treating this as a consumer mental-health risk in private life, not workplace crisis intervention.
OpenAI's official line is that no system is perfect, and that every notification is reviewed by trained people before it is sent, with a goal of finishing that review within one hour.
Does This Solve the Problem?
Partly.
It connects a broken chain in which AI can detect risk but has no downstream action. But it leaves several hard problems untouched:
- False positives: suicide-risk detection is not a binary classification problem, and false positives may be common. A person writing fiction, conducting academic research, or venting could trigger a review. One mistaken Trusted Contact activation can turn a nonexistent crisis into a disclosure inside a real relationship. OpenAI is relying on human review and a one-hour SLA to manage that balance, but nobody knows whether it will hold at scale.
- Misuse: it is easy to imagine a teenager-parent scenario in which a parent pressures the child to point Trusted Contact at them, creating a low-friction monitoring channel. OpenAI's defense is that the user must set it up themselves, but in practice there is little to stop a parent from sitting at the child's computer and clicking through the settings.
- It is not professional care: Trusted Contact only passes a risk signal to someone in the user's real life. What that person does next, whether they are able to help, and whether they respond professionally are all outside OpenAI's control.
Still, the feature does not need to solve everything. Its real purpose is to address the legitimacy of an AI company profiting from conversation traffic while leaving users' mental state entirely to an algorithm.
Giving the user control over who may be notified, giving trained humans the review role, and pushing the actual intervention back into real relationships are the parts Trusted Contact gets right.
It is also one of the first moments in the past two years when an AI company has been willing to admit that 1 million people talking to AI about suicide is not an engineering problem. It is a social problem.
Sources: CocoLoop, Introducing Trusted Contact in ChatGPT (OpenAI); ChatGPT can reach out to a friend if you're at risk of self-harm (Engadget); OpenAI Adds Trusted Contact Safety Feature to ChatGPT for Self-Harm Risk Notifications (gHacks Tech News)