openaienmodel: gpt-5-mini-2025-08-07
Introducing Trusted Contact — opt-in safety notifications in ChatGPT
Key Points
- Opt-in Trusted Contact for adults
- Human-reviewed alerts with sub-hour goal
- Limited notifications — no chat transcripts
Summary
Trusted Contact is an optional, opt-in safety feature that lets adult ChatGPT users nominate a trusted person (friend, family, caregiver) to be notified if automated detection and trained reviewers judge a conversation to indicate a serious self-harm concern. The feature supplements localized crisis resources and parental-control alerts, preserves user privacy, and is designed to encourage connection to real-world support rather than replace professional care.
Key Points
- Opt-in enrollment: users add one adult Trusted Contact (18+ globally; 19+ in South Korea) from settings. The contact must accept the invitation within one week.
- Detection and user prompt: automated systems detect potential self-harm risk and inform the user that a Trusted Contact may be notified; the user is encouraged to reach out with suggested conversation starters.
- Human review: a small team of specially trained reviewers examines flagged conversations before any notification is sent.
- Notification content & channels: if reviewers determine serious concern, a brief notification (general reason only, no chat transcripts or chat details) is sent via email, SMS, or in-app notification (if the contact has an account). The notification includes expert guidance links for checking in.
- Privacy and control: notifications are intentionally limited to protect user privacy; users can edit/remove their Trusted Contact at any time, and Trusted Contacts can opt out.
- SLA and scope: safety notifications are rare; OpenAI aims to complete human review and send notifications in under one hour when needed.
- Clinician guidance: feature was developed with clinicians, researchers, and organizations (e.g., APA) and complements existing crisis hotline guidance and parental controls.
Implementation notes for engineers
- Flow components: opt-in enrollment, invitation acceptance, automated risk detection, user-facing prompt, queued human review, limited-notify dispatch (email/SMS/in-app), and audit/logging.
- Privacy constraints: do not include conversation content in notifications; log minimal metadata required for review and audit; ensure access controls for reviewer tools.
- Operational targets: support sub-hour review workflows; rate-limit and batching considerations for spike scenarios; international age thresholds and localized crisis resource integration.