OpenAI is adding a Trusted Contact feature to ChatGPT that allows users to designate a person to be alerted when the system detects serious safety concerns in a conversation, a first step toward embedding duty-of-care architecture into consumer AI that raises immediate questions about user consent, alert accuracy, liability, and what it means to deploy AI as quasi-clinical care infrastructure at scale.
The feature works through an opt-in flow where a user nominates a trusted contact, typically a friend or family member, who can receive an alert if ChatGPT detects signals consistent with crisis states: suicidal ideation, acute distress, or self-harm indicators. The system monitors conversational context using its persistent memory and safety classifier layers. When escalation thresholds are met, ChatGPT offers the user resources, encourages reaching out to the nominated contact, and in some configurations alerts the contact directly. Users retain control over who is designated, but the escalation logic itself is not fully user-configurable. OpenAI has not disclosed the precise trigger thresholds, which is both defensible from a safety standpoint and opaque from a user-trust perspective.
The precedent in telehealth is instructive. Telehealth platforms like Teladoc and MDLive use structured clinical protocols for safety escalation, with licensed clinicians making the call and clear legal frameworks defining duty of care. Social platforms like Instagram and TikTok have built detection systems that display crisis resources when searches or captions trigger safety classifiers. Neither is perfect. Instagram's suicide prevention prompts can be gamed or ignored. Telehealth escalation depends on the clinician's judgment in a single session. ChatGPT sits in a different category: it has persistent memory, longitudinal context, and millions of users sharing more with it than they tell their therapists. The opportunity and risk are both larger than any prior deployment.
For SF readers, Trusted Contact signals that consumer AI is leaving the chatbot category and entering something harder to define and govern. OpenAI is not calling ChatGPT a therapy product, which is legally correct. It is building safety features that function like clinical escalation workflows, which is practically identical to what mental health platforms do. That positioning creates a product advantage and a liability exposure simultaneously. The advantage: safety features become a trust moat. Users who need genuine support will prefer a platform that can escalate over one that terminates with a hotline number. The exposure: every misfire, a false positive alert that embarrasses or alarms the user, and every false negative, a crisis that the system missed, creates potential liability that OpenAI's current legal structure has not tested.
The emotional-support and companion AI startup category is watching this development closely. Woebot, Replika, and Hinge Health all built clinical safety escalation into their products as table stakes, not differentiators. They did it because regulatory and clinical advisory requirements demanded it. Consumer AI platforms that operate at scale without licensed clinical frameworks have taken different approaches: some embed hotline numbers in high-risk conversations, some terminate chats that mention self-harm, some route to human moderators. OpenAI's Trusted Contact feature does something none of those do: it loops in a person who knows the user. That personalisation is closer to how effective community mental health actually works, through relationships rather than crisis lines, and it is a significant design choice.
The privacy tension is real and will become more visible as the feature rolls out. Trusted Contact requires a user to consent to sharing their ChatGPT safety state with another person. That consent is explicit in the opt-in flow. The harder question is what happens when a user in acute distress did not set up a contact in advance, or when the designated contact is not reachable, or when the alert itself destabilises the relationship between the user and the contact. Safety system design in mental health is a field with decades of professional practice. OpenAI is entering it with a machine learning classifier and a product team, which is not the same thing.
Whether safety features become a competitive moat or regulatory necessity depends on how quickly regulators respond. The EU AI Act classifies AI systems used in health and emotional support contexts as high-risk, requiring conformity assessments and human oversight. US state-level telehealth laws govern who can provide mental health services and when escalation is mandatory. If ChatGPT's safety features are framed as clinical care, they invite regulatory scrutiny. If they are framed as voluntary user features analogous to Instagram's crisis resources, they may avoid it. OpenAI's framing as a safety feature rather than a therapeutic product suggests the latter strategy, but the deeper the feature goes into longitudinal crisis monitoring with real-world alert escalation, the harder that distinction is to maintain.
Also read: DeepMind's AlphaEvolve uses Gemini to optimise its own infrastructure and rediscover mathematics • Microsoft's VS Code Copilot co-author blunder reveals the governance tension in AI developer tools • OpenAI's MRC protocol makes cluster networking the latest frontier in the training efficiency race
