OpenAI introduces Trusted Contact: a human-reviewed safety feature for ChatGPT users at risk of self-harm

OpenAI on June 7, 2026 introduced Trusted Contact — an optional, opt-in safety feature in ChatGPT that can notify a designated contact if automated systems and trained human reviewers determine a user may be at serious risk of self-harm. The feature is a direct response to longstanding concerns about AI chatbots interacting with vulnerable users in crisis.

What's new

Trusted Contact works as follows:

• User setup: Adults (18+ globally; 19+ in South Korea) can designate one trusted person — a friend, family member, or caregiver — from their ChatGPT settings. The designated contact must accept within one week for the feature to activate.
• Two-stage detection: If ChatGPT detects a conversation involving potential self-harm, the flagged conversation is reviewed by "a small team of specially trained people." Only if those human reviewers confirm a serious safety concern does a notification go out.
• Notification format: The trusted contact receives a brief notification (email, text, or in-app) indicating the general concern — that self-harm came up in a potentially concerning way. Chat transcripts are not shared.
• Privacy design: User conversations are not visible to the trusted contact; only the general concern flag is communicated.
• Opt-in only: The feature is disabled by default; users activate it voluntarily.

Context

AI chatbots have faced serious scrutiny over their handling of users in mental health crises. A Florida lawsuit filed in 2026 links ChatGPT to violent incidents, and multiple investigations have documented cases where AI chatbots responded poorly to users expressing suicidal ideation.

The challenge is structural: AI assistants are available around the clock and are often perceived by users as non-judgmental listeners. That accessibility makes them a de facto first contact for some users in crisis — a role they are not designed or equipped to fill.

OpenAI's approach borrows from established crisis intervention frameworks: connecting at-risk individuals to their existing human support networks, rather than having the AI act as the sole responder.

Why it matters

Trusted Contact is notable for several design choices that reflect lessons from prior mental health AI controversies.

First, the human review layer is significant. The feature does not fully automate crisis detection — it routes flagged conversations through trained reviewers before any notification is sent. This reduces false positives and keeps humans in the loop for decisions that affect vulnerable people.

Second, the transcript privacy commitment addresses one of the most sensitive concerns in this space: that AI safety features could become surveillance mechanisms. By limiting the notification to a general concern flag — without sharing conversation content — OpenAI preserves user privacy while still enabling safety intervention.

Third, the opt-in design respects user autonomy while making the safety infrastructure available to those who want it.

For AI policy, features like Trusted Contact are shaping what "adequate safeguards" looks like in regulatory discussions about AI and mental health. The EU AI Act and U.S. state legislators have been increasingly focused on this area.