OpenAI is rolling out a new optional safety feature for ChatGPT that aims to do something many people have long wished AI systems could do better: intervene earlier, and in a way that connects someone in distress to real-world support.
The feature is called “Trusted Contact.” In practice, it lets adult users designate a person they trust—such as a friend, family member, or caregiver—to be notified if OpenAI detects that the user may be discussing topics related to self-harm or suicide. The notification is intended to act as an additional layer of help alongside existing crisis resources, including localized helplines already available through ChatGPT’s safety tooling.
This is not framed by OpenAI as a replacement for emergency services or hotline infrastructure. Instead, it’s positioned as a bridge between what an AI can detect in a conversation and what humans can do when someone is at risk. The underlying idea is straightforward but significant: when a person may be in crisis, reaching out to someone they know and trust can make a meaningful difference—especially when that person might not otherwise seek help on their own.
What makes this announcement notable isn’t only the existence of another safety mechanism. It’s the direction of travel. For years, AI safety efforts have largely focused on preventing harmful outputs, filtering certain categories of content, or providing users with links to resources. Trusted Contact shifts part of the safety model from “information delivery” to “support escalation,” where the system can involve a third party under defined conditions.
OpenAI’s description emphasizes that the feature is optional and designed around a simple premise validated by experts: in moments of crisis, social connection matters. That premise is widely supported in mental health research and crisis intervention practice. People are more likely to accept help when it comes from someone they trust, and those trusted relationships can reduce isolation—the factor that often intensifies risk.
Still, the concept raises immediate questions that responsible AI observers will want answered clearly: What exactly triggers a notification? How does OpenAI determine that a conversation indicates potential self-harm or suicide risk? How much context is considered? What safeguards exist to prevent false alarms? And how does the system handle privacy, consent, and the emotional consequences of contacting someone who may not have been expecting such news?
OpenAI’s announcement, as reported, focuses on the high-level function: adults can assign an emergency contact for mental health and safety concerns, and that contact may be notified if OpenAI detects that the user may have discussed topics like self-harm or suicide with the chatbot. The feature is described as “another layer of support” alongside localized helplines already available. In other words, the system is meant to complement existing pathways rather than create a parallel crisis response that could confuse users about where to go for immediate help.
But even with that framing, the mechanics matter. Trusted Contact implies a detection pipeline that can recognize when a user’s language, intent, or pattern of conversation suggests heightened risk. That detection is likely built on a combination of content classification and contextual analysis—because simply matching keywords is rarely enough to distinguish between someone talking hypothetically, expressing intrusive thoughts without intent, or describing past experiences. The difference between “I’m thinking about hurting myself” and “I’m writing a story about self-harm” can be subtle, and the cost of getting it wrong is high.
If the system is too sensitive, it risks notifying loved ones unnecessarily, potentially causing distress, conflict, or a sense of betrayal. If it’s too conservative, it may miss moments when a user is silently signaling that they need help. Trusted Contact therefore sits at the intersection of safety engineering and human impact. It’s not just a technical feature; it’s a social intervention.
That’s why the “optional” nature of the feature is crucial. Users who opt in are effectively agreeing to a specific kind of escalation. They’re also choosing who will be contacted, which can reduce uncertainty and improve the odds that the notification reaches someone capable of responding appropriately. A trusted contact is not just any number in a database—it’s a person the user has selected. That selection process is part of the consent architecture: it gives users some control over who gets pulled into a crisis moment.
At the same time, opting in doesn’t eliminate the ethical tension. Even with consent, there’s a difference between agreeing to be helped and having your private conversation become a signal that triggers outreach to someone else. For many users, especially those who have experienced stigma around mental health, the fear isn’t only about being judged—it’s about being exposed. Trusted Contact attempts to address that by making the feature opt-in, but the emotional reality remains: the user’s words can have consequences beyond the chat window.
OpenAI’s positioning suggests the company is trying to balance two competing goals: protecting user privacy while also ensuring that safety interventions are effective when a user may not be able to act quickly. In crisis situations, people often don’t reach out for help in time. They may be overwhelmed, ashamed, or simply unable to navigate the steps required to get immediate support. An AI system that can detect risk and then connect the user to a trusted person could reduce the time between “I’m not okay” and “someone is checking on me.”
There’s also a practical dimension. Localized helplines are valuable, but they require the user to click, read, and decide to call or text. Some users may not have the energy to do that. Others may not know which resource fits their location. Trusted Contact adds a different pathway: instead of asking the user to take action, it asks the system to mobilize a human relationship.
This is where the feature becomes more than a safety toggle. It reflects a broader shift in how AI platforms think about responsibility. Safety features are increasingly moving toward “systems of care,” where the AI doesn’t just warn or redirect—it coordinates. That coordination can include providing resources, encouraging professional help, and now, potentially, involving a trusted person.
However, coordination introduces complexity. A trusted contact notification must be handled carefully to avoid escalating harm. If the message is too vague, it may not prompt action. If it’s too detailed, it could violate privacy or reveal sensitive information unnecessarily. The design of the notification content—what it says, what it doesn’t say, and how it frames the situation—will likely determine whether Trusted Contact helps or harms.
Another question is timing. Crisis detection is rarely instantaneous. Risk signals can build over time, and conversations can change tone quickly. If notifications are sent too early, they may interrupt a user who is seeking support but not in immediate danger. If notifications are delayed, the opportunity to intervene may be lost. OpenAI’s announcement doesn’t specify timing details, but the feature’s effectiveness depends heavily on how quickly the system decides to escalate.
Then there’s the issue of jurisdiction and emergency response. In many regions, contacting a trusted person is not the same as contacting emergency services. A loved one may not be able to provide immediate medical intervention. But they can often do what AI cannot: check on the person, encourage them to seek help, stay with them, and call emergency services if needed. Trusted Contact can therefore be seen as a “first human responder” step—one that can lead to further escalation when appropriate.
This is also why the feature is framed around mental health and safety concerns rather than a general “any harmful topic” policy. The scope matters. By focusing on self-harm and suicide-related discussions, OpenAI is targeting a category where the stakes are highest and where timely human support can be life-saving. At the same time, limiting the feature to specific risk domains may reduce the chance of unnecessary notifications for less severe issues.
For users, the decision to opt in may depend on their personal circumstances. Someone living alone might benefit more from a trusted contact being alerted, because they may not have anyone nearby to notice changes. Someone with a supportive network might feel comfortable selecting a trusted person. But someone whose relationships are strained—or who fears that a notification could trigger conflict—might choose not to enable the feature.
That variability is important. Responsible safety design often means giving users meaningful choices rather than forcing a one-size-fits-all approach. Trusted Contact being optional suggests OpenAI is acknowledging that users have different preferences and different risk tolerances.
From a broader perspective, the feature also invites scrutiny about how AI companies define “detection.” When OpenAI says it “detects” that a user may be discussing self-harm or suicide topics, it implies a threshold. Thresholds are where safety systems become either protective or problematic. A threshold that is too low can create a flood of alerts; a threshold that is too high can fail to catch genuine crises. The company’s ability to calibrate that threshold—using data, evaluation, and ongoing monitoring—will determine whether Trusted Contact becomes a reliable safety net or a source of anxiety.
There’s also the question of how the system learns and improves. Safety features typically evolve based on feedback loops: user reports, internal audits, and performance metrics. But mental health risk detection is particularly sensitive to bias. Language patterns vary across cultures, age groups, and communities. Some users may express distress indirectly. Others may use slang or coded language. If the detection model isn’t robust across these variations, Trusted Contact could disproportionately misread certain users’ conversations.
OpenAI’s announcement references expert validation of the premise, but the technical validation of detection accuracy is equally important. Experts can validate the concept of connecting to trusted people, but the system still needs to correctly interpret the conversation signals that precede crisis. That’s where transparency and evaluation matter.
Even so, the feature’s existence is a sign of how quickly AI safety expectations are changing. Users increasingly expect AI systems to do more than respond politely. They want AI to recognize when a conversation is dangerous and to take action that aligns with real-world safety practices. Trusted Contact is one attempt to meet that expectation.
It also reflects a growing recognition that mental health support is not purely informational. It’s relational. A chatbot can offer empathy, coping strategies, and resources, but it cannot replace the human presence that often makes the difference during