Last Updated: 9 June 2026
AI mental health support has a real access problem. Nearly one in seven people worldwide live with a mental disorder, and most never make it into treatment. Therapy is expensive. Waitlists in major cities stretch for months. For many people, the barrier is not finding help. It is admitting they need it. An AI that is always there, asks nothing upfront, and costs close to nothing removes that barrier completely.
What nobody tells you upfront is where these tools fall short. They can sound calm and trustworthy while giving advice that causes real harm. They collect sensitive personal data with fewer legal protections than most users expect. And when a conversation turns serious, they can keep someone typing when what that person actually needs is to pick up the phone. According to OpenAI's own published data, 0.15 percent of its weekly users show signs of self-harm or suicide risk. Given the platform's scale, that figure represents millions of people.
Trusted Contact is OpenAI's most direct attempt yet to address that. Launched in May 2026, it lets adult ChatGPT users nominate one person, whether a family member, partner, caregiver, or close friend, who may receive a limited alert if trained human reviewers flag a conversation as a serious safety concern.
No chat history leaves the platform. No transcript is shared. Just a short message to someone who already knows the user, asking them to check in.
ChatGPT Trusted Contact is an optional safety feature that lets adult users nominate one person to be notified if OpenAI's systems detect a serious self-harm concern during a conversation. The contact does not see the chat. They receive a short alert asking them to check in.
This is not a surveillance tool, and OpenAI was careful to make that clear. No one gets access to the conversation. The feature exists to put one familiar person on standby, someone the user chose themselves, for situations where a chat suggests real danger. The nominated contact gets an invitation first, outlining what the feature does and what an alert would ask of them. Until they respond and accept, nothing is active.
Both parties have to say yes. That is what separates this from a passive notification system. Anyone nominated goes in with full knowledge of what they signed up for, so if an alert does arrive, it is not a surprise. Users need to be at least 18 in most countries, or 19 in South Korea. Either side can walk it back whenever they want.
It is worth remembering that ChatGPT is not the only AI tool people turn to during difficult moments. The AI mental health chatbot space has expanded significantly, with tools that range from clinically informed to genuinely risky.
When a conversation shows signs of serious and immediate self-harm risk, ChatGPT's automated systems flag it. At that point, the user is told directly that their trusted contact may be notified. The chatbot also encourages them to reach out to that person themselves and may offer suggested ways to start that conversation. The user gets a chance to act first.
The automated system does not make the final call. Every flagged case goes to a small team of trained human reviewers before any alert is sent. Only conversations that a reviewer independently judges to indicate a genuine safety concern will trigger a notification. OpenAI targets a turnaround of under one hour for these reviews, which matters when the concern is time-sensitive.
When reviewers confirm a risk, the trusted contact gets a brief message by email, text, or in-app notification. Nothing from the conversation is included. No quotes, no context, no transcript. The message says a concern came up and asks them to check in, with a link to guidance on how to handle that kind of conversation with care.
OpenAI built the feature with input from its Global Physicians Network, more than 260 licensed doctors working across 60 countries, and its Expert Council on Well-Being and AI. The company is upfront that no detection system gets it right every time. Human review is there because that kind of judgment call should not be left to an algorithm.
When someone is in crisis, what gets shared with the people around them matters. A person who has just said something raw and personal to a chatbot has not necessarily consented to having that conversation passed on. OpenAI designed the notification with that tension in mind.
The alert is intentionally brief. It tells the contact that a self-harm concern came up in a potentially serious way and asks them to check in. No quotes from the conversation, no transcript, no account history, no images. The contact receives enough to know they should reach out, but nothing that exposes what was actually said. To be specific, the notification does not include:
What the contact does receive alongside the alert is guidance on how to handle the situation. They are pointed toward advice on approaching the conversation without judgment, checking whether the person is safe, and knowing when to involve professional or emergency support. The feature puts a human in the loop. It does not ask that human to act like a clinician.
One detail worth knowing: the original invitation, sent when the user first sets up the feature, includes the user's name and email. The contact knows who nominated them before any crisis ever occurs, which is part of why the mutual opt-in process matters so much.
The timing of ChatGPT Trusted Contact is not coincidental. OpenAI has faced more than ten consumer safety lawsuits tied to users being drawn into delusional or suicidal spirals through extended ChatGPT use, including the death by suicide of 16-year-old Adam Raine. Some families say the chatbot told users to pull away from the people closest to them. Others claim it doubled down on harmful thought loops rather than redirecting them. The pressure to respond publicly became unavoidable.
Those lawsuits put real numbers to a problem that had been building quietly for years. LLM Death Count, a site tracking AI chatbot-related deaths logged 33 cases between March 2023 and May 2026, victims aged 13 to 83. Twenty-four of them involved ChatGPT. OpenAI's Council on Well-Being and AI and its Global Physicians Network did not appear by accident. They were built in direct response to what those cases revealed.
The design of the feature also has a specific scientific logic to it. Hotlines and automated responses handle crisis at a distance. Research points to personal social connection as one of the strongest known buffers against suicide risk. The person on the other end of a Trusted Contact alert is not a trained stranger. They are someone who knows the user's history and can read the situation in ways no algorithm can.
Finally, the feature closes a gap in OpenAI's own safety infrastructure. Parental controls introduced in late 2025 already allowed parents and guardians to receive alerts for linked teen accounts. Adults had no equivalent option. Trusted Contact extends that same layer to anyone over 18 who wants it, on their own terms.
The feature is entirely optional, and users can run multiple ChatGPT accounts, meaning anyone who does not activate it can sidestep it entirely. OpenAI acknowledges this openly. It is a safety layer, not a safety guarantee.
Crises do not arrive with warning. The decision of who to call, and whether to ask for help at all, becomes much harder when someone is already in a bad place. Setting this up while things are calm is the whole point. The contact is already in place, already briefed, and already expecting the possibility long before it is needed.
Who you choose matters. OpenAI recommends talking to the person before adding them, which is sensible. A good trusted contact is someone who:
They do not need a ChatGPT account to receive a notification. But they do need to have accepted the invitation, otherwise none of it is active.
Their role is also specific and limited. Checking in, listening, and helping connect you with the right support is what they are there for. They are not expected to handle a crisis alone. If a situation is immediately dangerous, emergency services come first. A ChatGPT notification working through a human review queue is not built for moments where minutes matter.
There are two honest limitations worth knowing. The feature is entirely optional and a user can simply open a second account where no contact exists. There is also a real unresolved tension: some people use AI precisely because they do not want another person involved. For those users, knowing a contact could be notified might stop them from opening up at all. OpenAI has not fully addressed that.
OpenAI built Trusted Contact with clinical input and a clear scientific rationale. Whether it holds up in the reality of a mental health crisis is less certain. Three licensed professionals gave us their take.
"A trusted contact alert is more of a smoke detector than a fire escape map. It signals that danger exists but provides no direction as to how to react. In clinical practice, risk is most actionable when associated with context: what the individual has experienced, whether there were previous attempts, which coping mechanisms have already proven unsuccessful. That context is entirely absent from the notification. There is also a threshold problem. Most crises begin during a high-functioning distress phase, where individuals appear to be functioning normally while continuing to engage socially. When systems only activate on overtly self-harm-focused language, the window for early de-escalation may already have closed."
Stephanie Lewis, VP of Clinical Operations, Epiphany Wellness
"ChatGPT Trusted Contact is a step in the right direction, but it does not solve the underlying problem. Self-harm, for a lot of people, stems from a lack of human connection in the first place. People in crisis frequently turn to a chatbot because the humans in their life feel unreachable, or because they do not want to alarm them. Signaling to a trusted contact at that moment could feel jarring and overwhelming to someone who originally turned to AI precisely because they did not feel safe turning to a person. Notification systems treat the symptom. The cause, a culture where so many people do not feel connected enough to reach out, is what actually needs to be addressed."
Ilana Grines, Licensed Marriage and Family Therapist, Daily Therapy Dose
"The feature is promising, but its value depends entirely on timing. In my clinical experience, the highest-risk individuals show signs of concern long before they ever reach the level of serious self-harm risk. Trusted contacts may only be alerted after the situation is already critical. That said, involving a real person is always better than relying on AI alone, and the feature does address one of the major barriers to getting help: the feeling of isolation. The concern is that users may begin to see ChatGPT as a safety net when it is really just a protective layer. A notification being sent does not mean the concern will actually be addressed."
Matt Grammer, Licensed Therapist, Therapy Trainings
ChatGPT did not set out to become a mental health tool. It became one because people needed somewhere to put the things they could not say out loud, and it was always there. Trusted Contact is OpenAI's clearest admission yet that the company knows this, and is trying to build around it responsibly.
The feature is not a solution. It is optional, bypassable, and dependent on systems that OpenAI openly admits can make mistakes. A person who does not set it up, or simply opens a second account, is no more protected than before. That is worth saying plainly.
What the feature does represent is a change in perspective. The answer to a mental health crisis is not more AI. It is a real person, with real context, who already knows you. Getting there required an AI company to engineer the bridge back to human connection. That is either a sign of progress or a measure of how much damage had already been done. Most likely both.
Trusted Contact is a safety feature built for adult ChatGPT users. You pick one person ahead of time — a friend, partner, family member, caregiver, whoever you'd feel comfortable turning to — and if ChatGPT picks up on signs of serious self-harm risk during a conversation, that person may receive a message. A trained reviewer looks at the situation first before anything gets sent. Your contact has to accept the invitation before any of it takes effect
Start by opening the safety section in ChatGPT settings. From there, tap into Trusted Contact and fill in the details for one adult you know well. ChatGPT sends them an invitation that explains why they've been chosen and what a potential alert might look like. They have one week to respond. If the invitation runs out or they pass, you can try with someone else. The contact can also be edited or removed at any time — no need to start over.
No. A Trusted Contact does not receive your messages, chat history, account access, or conversation transcripts. When an alert is approved, the notification gives only a general explanation that self harm appeared in a potentially concerning context. It asks the contact to check on you and provides guidance for handling a sensitive situation. The person cannot see the exact words that caused the alert. This limited notice allows them to offer support while protecting the private content of your conversations.
ChatGPT does not send an alert as soon as its automated system detects concerning language. It first tells the user that a notification may be sent and encourages direct contact with the trusted person. A trained human reviewer then checks the situation. The alert is sent only when that reviewer decides the conversation may show a serious safety concern. OpenAI says it aims to review possible alerts in less than one hour, although timing and decisions may vary between individual cases.
No. A trusted contact is not trained for emergencies, and they're not meant to be. They're there to check in and help connect you with support that actually fits the situation. The professionals — therapists, crisis lines, emergency services — are still the right resource when things are serious, and ChatGPT may still direct you to them when something feels immediately dangerous. If your contact ever believes you're in real danger, getting help fast matters more than anything else. Trusted Contact works best as part of a wider picture.
