Jul, 134 min to read
Frontend
News · OpenAI adds Trusted Contact, an opt-in crisis-notification flow, to ChatGPT
Jul, 94 min to read
FrontendOpenAI adds Trusted Contact, an opt-in crisis-notification flow, to ChatGPT
A new setting lets adults nominate someone who may be alerted if trained reviewers judge a conversation to signal serious self-harm risk — with a deliberately thin notification payload.
What the setting actually adds to the app
Trusted Contact is an optional feature configured from ChatGPT settings. An adult user (18+ globally, 19+ in South Korea) can nominate one person — a friend, family member, or caregiver — who may be notified if OpenAI's automated systems and trained reviewers detect that the enrolled user may have discussed harming themselves in a way indicating a serious safety concern.
OpenAI frames it as an additional layer alongside the localized crisis helplines already surfaced in ChatGPT, not a replacement for professional care. The feature extends existing parental-control safety notifications for linked teen accounts to any adult who opts in.
The enrollment flow is built around consent and revocation
The frontend mechanics are specific. A user adds exactly one adult contact. That contact receives an invitation explaining their role and must accept within one week for the feature to activate. If they decline, the user can nominate someone else.
Both sides retain control: users can edit or remove the contact in settings, and the contact can remove themselves at any time via the help center. This two-sided opt-in — the enrolled user chooses, and the nominated person must actively accept — is the load-bearing design decision here, and it distinguishes the feature from a silent monitoring alert.
The notification is intentionally thin
When automated monitoring flags a possible self-harm concern, ChatGPT first tells the user it may notify their Trusted Contact and encourages the user to reach out themselves, offering suggested conversation starters. A small team of specially trained reviewers then evaluates the situation before anything is sent.
If reviewers confirm a serious concern, the contact gets a brief notification by email, text, or in-app message. Critically, that message shares only the general reason self-harm came up and a link to expert guidance — no chat details or transcripts. The payload is deliberately limited to protect the enrolled user's privacy while still prompting a real-world check-in.
A human reviewer sits in the loop, with a stated latency target
OpenAI states that every notification undergoes trained human review before sending, and that it strives to review these safety notifications in under one hour. That is a concrete operational commitment: the automated classifier is a trigger, not the decision-maker, and there is a person and a time budget between detection and alert.
OpenAI also acknowledges the failure modes plainly — no system is perfect, and a notification may not always reflect what someone is actually experiencing. The human review step is positioned as the mitigation for false signals.
The implication: crisis response as a product surface, not just a model behavior
What OpenAI shipped here is mostly frontend and workflow: a settings entry, an invitation-and-acceptance handshake, an in-conversation prompt, a reviewer queue with a latency goal, and a privacy-constrained message template. The underlying detection and refusal behaviors already existed; Trusted Contact routes a confirmed signal to a specific human the user pre-chose.
The design bet, backed by the APA quote on social connection as a protective factor, is that the most useful move in a crisis is not a better model response but a bridge to a named real person. For teams building sensitive-conversation interfaces, the notable pattern is the layering of consent gates and a human review step around an automated classifier — treating who gets told, how fast, and how little to say as first-class product decisions.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.