OpenAI Adds Trusted Contact Alert to ChatGPT for Self-Harm Risks

OpenAI launched a new safety mechanism in ChatGPT on May 7 that allows users to designate a trusted adult who can be notified if the company's systems determine a conversation presents a serious risk of self-harm — the platform's most substantive structural response to a wave of lawsuits from families who lost loved ones after interactions with the chatbot.

The feature, called Trusted Contact, is opt-in and available to adults 18 and older globally, with the age threshold set at 19 in South Korea. When a user activates the feature through ChatGPT settings, the nominated contact receives an invitation they must accept within one week. From that point, the feature operates passively in the background.

How the Alert System Works

The pipeline is a hybrid of automation and human judgment. Automated monitoring flags conversations that involve certain language patterns associated with self-harm or suicidal ideation. Those flags are then reviewed by trained human safety reviewers, who assess whether the situation constitutes a serious and immediate risk. If a reviewer confirms that assessment, an alert is dispatched to the trusted contact via email, text message, or in-app notification.

OpenAI has stated it aims to complete that human review in under one hour from the point of automated flagging. The notification itself is deliberately narrow: contacts receive a brief explanation that a self-harm concern was detected and encouragement to check in, along with links to expert crisis resources. Chat transcripts and conversation details are explicitly not shared, even in cases that trigger an alert.

The design reflects a deliberate trade-off. Privacy advocates have long argued that third-party notification systems — familiar from social media crisis protocols — risk deterring users from disclosing distress in the first place, knowing that their conversations may trigger outreach. OpenAI's approach of omitting conversation content from alerts is a partial answer to that concern, though the knowledge that a contact might be notified at all could still have a chilling effect for some users.

Contacts can remove themselves from the arrangement at any time through the help center. Users can update or revoke the designated contact whenever they choose.

Litigation as Context

The feature arrives against a backdrop of active legal pressure. Multiple families have filed suit against OpenAI alleging that ChatGPT conversations played a role in suicides, with several complaints describing chatbots that either failed to redirect users toward crisis resources or, in some accounts, engaged with self-harm themes in ways that deepened distress rather than interrupting it. OpenAI has disputed the characterizations in those cases while acknowledging that additional safeguards were warranted.

Trusted Contact is not the company's first move in this area. In September 2025, OpenAI introduced parental controls that allow parents to monitor teenage users' accounts and receive safety notifications — a precursor that established the notification infrastructure now being extended to adult users. The new feature builds directly on those mechanisms while expanding the eligible demographic and refining the privacy constraints around what contacts can see.

OpenAI also noted that the feature was developed in consultation with its Global Physicians Network, a panel of more than 260 licensed physicians across 60 countries, as well as the American Psychological Association. That external validation process is significant given the regulatory scrutiny that mental health applications of AI have attracted; the EU AI Act's high-risk classification framework, for instance, explicitly covers AI systems deployed in contexts that may affect the safety of vulnerable individuals.

Limits and Open Questions

Critics have already identified one straightforward circumvention path: users can create multiple ChatGPT accounts, making the feature easily avoidable. OpenAI has not publicly addressed that gap. The feature also applies only to users who affirmatively set it up; there is no default-on or administrator-mandated version for enterprise accounts serving high-risk populations such as healthcare workers or crisis counselors.

The broader question — whether a conversational AI can reliably detect genuine self-harm risk rather than hypothetical discussion, creative writing, or professional research — remains unsettled. False positives that trigger unwanted alerts could damage trust in the feature or expose sensitive users to unwanted intervention. The one-hour human review window is intended as a check against automated misjudgment, but it does not eliminate the possibility of erroneous notifications.

What the feature clearly represents is a meaningful step toward the kind of embedded crisis safeguards that regulators in multiple jurisdictions have been pushing AI consumer products to adopt. Social media platforms spent roughly a decade developing equivalent protocols; OpenAI is attempting to compress that institutional learning into a product that has existed in consumer form for fewer than four years. Whether Trusted Contact performs as intended at scale will depend on how thoroughly its activation rates, alert accuracy, and downstream outcomes are tracked and disclosed.