What happens when someone using ChatGPT shows signs of a mental health crisis and there's no one else in the room? OpenAI's answer is Trusted Contact - an optional feature that routes a human into the loop when the model detects serious self-harm concerns.
The setup is straightforward: users add a trusted person's contact details inside ChatGPT. If the app detects signs of serious self-harm risk during a conversation, that contact gets notified. The feature is opt-in only, and OpenAI announced it on May 7, 2026.
This is a meaningful design shift from how AI safety typically works. Most guardrails operate silently - blocking certain outputs or appending crisis hotline numbers at the bottom of responses. Trusted Contact does something more concrete: it creates a direct connection between an AI conversation and someone in the physical world.
The population this helps most is people who want a backstop but find it hard to ask for one directly. Setting up a trusted contact is a lower-friction way of signaling "I might need someone to check on me" without having to say those words out loud.
The practical details matter a lot, though. OpenAI describes the trigger condition as "serious self-harm concerns" - deliberately broad. The detection threshold will determine whether this works as intended. Set too sensitive, and it fires on false positives, eroding trust and potentially straining relationships. Set too conservative, and it misses what it was built for.
There's also a privacy tradeoff worth understanding before opting in. Enabling this feature means a third party could receive a notification based on conversation content. That's not a flaw in the design, but it should be front of mind when configuring it.
How well the detection holds up in practice is genuinely unknown. AI models have a mixed track record in mental health contexts - sometimes reading distress accurately, sometimes missing signals, sometimes triggering on benign text. The entire value of Trusted Contact rests on whether that underlying detection is reliable enough to justify the trust users place in it.