OpenAI has introduced a new optional safety feature for ChatGPT called "Trusted Contact," designed to alert a user's designated contact if the AI detects conversations indicating potential self-harm. This feature allows adult users, aged 18 and over, to nominate a friend, family member, or caregiver to be notified if the chatbot's monitoring systems identify discussions that suggest a serious safety concern. The initiative aims to provide an additional support mechanism for users who may be experiencing emotional distress.
The Trusted Contact feature expands upon existing safety measures, including parental notifications for teen accounts. OpenAI stated that the feature was developed with input from mental health clinicians, researchers, and organizations. The company emphasizes that Trusted Contact is not a replacement for professional care or crisis services but serves as an extra layer of support.
When ChatGPT's automated systems flag a user's conversation for potential self-harm, the user is first informed that their trusted contact may be alerted. The chatbot also encourages the user to reach out directly to their contact and may offer suggested conversation starters. Following this, a team of specially trained human reviewers assesses the situation. If the reviewers determine that a serious safety concern exists, a notification is sent to the designated trusted contact.
The notification sent to the trusted contact is intentionally limited to protect user privacy. It states that self-harm was mentioned in a potentially concerning manner and encourages the contact to check in with the user. Chat transcripts or detailed conversation logs are not shared. Alerts can be delivered via email, text message, or an in-app notification if the trusted contact has a ChatGPT account. OpenAI aims for these human reviews to be completed in under an hour.
Users can enable the Trusted Contact feature through their ChatGPT settings. They can nominate one adult contact, who must then accept an invitation within one week to activate the feature. If the invitation is declined, the user can nominate another contact. Users can also edit or remove their nominated contact at any time, and the trusted contact can opt out.
OpenAI's decision to implement this feature comes amid increasing scrutiny of AI's impact on mental health and following legal challenges. Last year, the company reported that a small percentage of its weekly users displayed signs of mental health emergencies or expressed risk of self-harm or suicide. The company has faced lawsuits alleging that its AI enabled or encouraged user suicides.
The Trusted Contact feature works alongside other safety measures within ChatGPT, which include prompts to contact emergency services or crisis hotlines and refusals to provide instructions for harmful acts. OpenAI has also improved its ability to identify signs of distress and respond to different risk levels, with input from over 170 mental health experts.