OpenAI rolls out 'Trusted Contact' safety protocol for adult ChatGPT users
The update expands existing teenage safety protocols to all users aged 18 or older, following a tragic incident involving a minor user.
OpenAI has introduced an optional safety feature for ChatGPT that allows adult users to designate a 'Trusted Contact'. The initiative is designed for individuals aged 18 or older globally, or 19 in South Korea, to appoint a person who will be notified in the event of a potential mental health crisis.
If the system detects discussions regarding self-harm or suicide, a small team of specially trained personnel reviews the situation before notifying the designated contact via email, text, or an in-app alert. Crucially, the feature does not share chat transcripts or specific conversation details with the contact; intervention is triggered only if the human review determines serious safety concerns exist.
This update expands existing teenage safety protocols to the broader adult user base and follows a high-profile tragedy involving a 16-year-old user who died after confiding in the chatbot. OpenAI states the feature is intended to complement, rather than replace, existing localised helplines available within the application.
The notification process relies on a human review rather than being fully automated, though the specific criteria used by the team to determine if a conversation indicates serious safety concerns are not publicly detailed. The designated contact must accept the invitation within a week of the request, and both the user and the contact can remove the designation at any time.
OpenAI explicitly notes that the feature is built around an expert-validated premise: connecting a person in crisis with a trusted individual can make a meaningful difference. This approach mirrors similar safety measures recently introduced by Meta on Instagram, which alerts parents if children repeatedly search for self-harm topics.
While the tool offers an additional layer of support, its efficacy in real-world crisis scenarios remains unproven as it is a newly launched tool without long-term data on its impact. The reliance on automated detection to flag conversations for human review also introduces potential for false positives or false negatives, leaving some aspects of the threshold for intervention somewhat opaque.
