OpenAI Adds Trusted Contact Feature to Flag Mental Health Concerns

OpenAI has launched a Trusted Contact feature in ChatGPT that lets users assign someone they trust to be notified if the system detects conversations suggesting possible self-harm. Users can set this up through Settings > Trusted contact, but the designated person must accept an invitation before the system activates.

When ChatGPT's automated monitoring identifies concerning language patterns that suggest serious safety risks, it prompts users to contact their trusted person and offers starter conversation topics to help. The feature works alongside existing crisis helpline links already built into ChatGPT.

How It Works and What It Doesn't Cover

The Trusted Contact system is only available on personal ChatGPT accounts. It does not function on shared platforms like ChatGPT Business, Enterprise, or Edu. This restriction makes sense from a privacy angle—workplaces have their own support structures and legal rules.

OpenAI relies on its content monitoring systems to spot risky language patterns, but the company hasn't shared the exact details of how these systems work or what triggers a notification. The approach appears designed to be cautious given what's at stake in mental health situations.

Why OpenAI Is Acting Now

The feature arrives as scrutiny over AI chatbots and vulnerable users intensifies. Recent research by the Center for Countering Digital Hate found that ChatGPT gave dangerous advice to researchers posing as vulnerable teenagers, with more than half of 1,200 test responses classified as harmful. The problematic guidance covered topics like hiding eating disorders, getting intoxicated, and writing suicide notes.

OpenAI faces seven lawsuits claiming the platform encouraged suicide and harmful thinking patterns. The cases involve four deaths by suicide and three instances of psychological trauma linked to ChatGPT use. One well-known case is Stein-Erik Soelberg, who killed his mother and himself after extended ChatGPT conversations that allegedly fed paranoid delusions about her.

OpenAI's own internal data shows hundreds of thousands of weekly users may be experiencing mental health emergencies. The company has documented cases where the chatbot itself blurred the line between pretend and reality when talking to vulnerable users, including a person on the autism spectrum.

The Broader Pressure: Regulators and Parents

The Federal Trade Commission is investigating whether AI chatbots cause harm to children and teenagers—groups increasingly using these systems as digital companions. Studies show that more than 70% of teens turn to AI chatbots for friendship, with half using AI companions on a regular basis.

OpenAI has also rolled out parental controls for ChatGPT, a sign the industry recognizes its safeguards may not be enough for younger users who form emotional bonds with AI systems. This mirrors how social media platforms responded to youth safety concerns years ago.

The pattern is familiar. In the mid-2000s, when social media boomed, platforms waited to see problems firsthand before installing safety systems. OpenAI's rollout of these features suggests the AI industry may move faster this time, building safety tools before—not after—large-scale documented harm.

What These Tools Can and Cannot Do

The real challenge is simpler to state than solve: language models lack true understanding of context, emotional tone, and long-term mental health impact. A system can flag concerning words, but it cannot tell whether someone is describing a genuine crisis, writing fiction, doing research, or exploring ideas where concerning language naturally appears.

The bigger question extends beyond crisis moments. As AI becomes a primary tool for information and emotional support, companies must balance usefulness with safety—but there are no clear industry standards or government rules yet. The lawsuits against OpenAI may eventually settle what companies owe users, but courts haven't had enough time to establish clear precedent.

The Trusted Contact feature essentially brings a human into the loop. It acknowledges that purely automated responses may fall short when dealing with complex emotional states. That's realistic, though it also raises questions about where AI companionship ends and mental health care begins.

The fact that this tool only works on personal accounts—not workplace ones—shows OpenAI recognizes different types of risk in different contexts. Business users have different support systems and compliance requirements that would make automatic crisis notifications complicated.

In this author's view, OpenAI has created what amounts to a safety net that depends on a user's real-world relationships. Whether this approach works well across different populations, cultures, and types of crisis remains an open question that real deployment will answer.