AI Counseling and Companion Era, Core Issue Moves to 'Human Connection' Beyond Technical Response
OpenAI has introduced a 'Trusted Contact' feature to ChatGPT. It is an opt-in safety feature that allows trusted people — family, friends, or guardians — pre-designated by the user to receive notifications in crisis situations.
This announcement is not simply a feature update. It is an important change emerging in an era when generative AI has come to face moments of emotional crisis and isolation, going beyond a tool that answers people's questions. This is close to a declaration that AI will not solve crises on behalf of people but rather become a path connecting to real people.
OpenAI announced on May 7, 2026 that Trusted Contact will be released sequentially. This feature is designed so that when an automated system and trained reviewers detect the possibility of serious self-harm concern, a limited notification can be sent to the designated contact. OpenAI's official blog and release notes also explained that this feature is sequentially provided for adult users of personal ChatGPT accounts and is not provided for Business, Enterprise, or Edu workspaces.
The 'Trusted Contact' Feature That Came Into ChatGPT
Trusted Contact works by the user designating one trusted adult in ChatGPT settings. In Korea, adults aged 19 or older, and globally adults aged 18 or older, can be registered as trusted contacts. The registered party receives an invitation and must accept within a week for the feature to be activated.
The core operation method is relatively clear. When a user shows serious warning signals related to self-harm, ChatGPT first informs the user that a notification may be sent to the trusted contact. Simultaneously, conversation starter phrases that the user can directly use to contact the person are also suggested.
Notifications do not immediately go out from the automated system alone afterward. A small number of separately trained personnel reviews the situation. If review results determine there is serious safety concern, a brief notification is sent to the trusted contact via email, text, or app notification.
An important point is that the scope of notification is limited. OpenAI explained that conversation transcripts or specific chat content are not included in the notification. Instead, general grounds that there was self-harm-related concern and guidance to check on the user are communicated.
The Moment AI Becomes a 'Connector' Rather Than 'Counselor'
The core of this feature is that AI does not try to solve crises alone.
Generative AI has already become a space for many people to share their worries. Users ask ChatGPT about not only studying, work, and creation but also loneliness, anxiety, relationship problems, and self-doubt. At this point AI can be an immediate, non-judgmental conversation partner. However, precisely for that reason danger also arises.
When a person is in serious emotional crisis, AI's answers alone are insufficient. What is needed is often real contact rather than accurate sentences. Someone calling, someone knocking on the door, someone being beside them.
What OpenAI emphasized in this announcement is precisely this point. Trusted Contact does not replace specialized treatment or crisis response services. Instead it is an additional safety layer that helps users connect with people they already trust in crisis moments.
That is, this feature shows the direction of AI safety. Not AI solving all problems itself but returning to people in necessary moments.
Combination of Automated Detection and Human Review
A noteworthy part of Trusted Contact's design is the combination of automated systems and human review.
OpenAI explains that automated monitoring systems can detect the possibility of serious self-harm concern in user conversations. However, notification sending is not determined by automated detection alone. Notification is sent to the trusted contact only when trained personnel reviews the situation and judges there is serious safety concern.
This shows the difficulties faced by AI safety systems. Self-harm warning signals are difficult to judge from a single sentence alone. Some expressions may be actual warning signals, but some may be metaphors, jokes, or past recollections. Conversely, people in dangerous states may avoid direct expressions.
Therefore simple keyword filtering alone is insufficient. Context must be read, risk levels must be distinguished, and balance between false positives and false negatives must be maintained. OpenAI's inclusion of a human review stage can be seen as a device to reduce such uncertainty.
However, OpenAI also acknowledges that the system is not perfect. Notifications may not always accurately reflect the user's actual state, but all notifications go through trained human review before sending.
The Boundary Between AI and Mental Health Is Blurring
Behind Trusted Contact's emergence is a change in how generative AI is used.
Initial ChatGPT was received as a productivity tool. A tool for writing emails, fixing code, organizing reports, and expanding ideas. However, over time people began using AI as a more private space. Emotions difficult to say, worries difficult to bring up even to family or friends, anxiety coming late at night — these are shared with AI.
This change is two-sided. AI can be the channel for someone's first request for help. However, simultaneously there is also the risk of AI becoming a space that delays real help.
Trusted Contact is a feature emerging from within this tension. The problem consciousness is that if AI is in an era where it must receive emotional conversations, safety devices are needed that connect to the real world when those conversations reach dangerous levels.
OpenAI stated that it reflected advice from clinicians, researchers, and mental health and suicide prevention organizations in developing this feature. It also explained that it collaborated with external organizations including a network of over 260 physicians in 60 countries, a wellbeing and AI expert committee, and the American Psychological Association.
Where Is the Balance Between Safety and Privacy Found?
As important as Trusted Contact is as a feature, controversy is also unavoidable. The biggest issue is the balance between safety and privacy.
On one hand it is clear. In situations where serious self-harm risk is detected and the user is isolated, trusted contact notification can help protect life. In particular, it can be difficult to request help oneself in crisis moments. If there is a pre-designated trusted contact, the possibility of the user not being left completely alone increases.
However, on the other hand sensitive questions remain. To what extent can AI judge danger? To what extent is user conversation monitored? How is user privacy protected when notifications are false positives? How can it be guaranteed that the trusted contact is truly a safe person?
OpenAI, apparently conscious of these concerns, designed the feature as opt-in. Users can delete or modify the trusted contact at any time, and the trusted contact can also remove themselves from the role. Specific conversation content or transcripts are not included in notifications.
Nevertheless, this feature demonstrates that AI services are increasingly entering more sensitive domains. Now AI safety is not simply a matter of "does it refuse dangerous answers." It is expanding to the problem of designing users' emotions, relationships, crisis signals, and real-world protection networks.
OpenAI's Safety Strategy Change: From Blocking Answers to Connecting Relationships
OpenAI stated it operates safety devices beyond Trusted Contact that guide connection to crisis counseling lines, emergency services, mental health professionals, and trusted people nearby in sensitive conversations. It also explained that requests asking about methods of suicide or self-harm are trained to refuse guidance and provide safer responses and region-specific crisis resources.
The change visible here is the expansion of safety strategy.
Past AI safety was primarily a blocking problem. Not answering dangerous questions, not providing illegal or harmful instructions, and preventing harmful content generation were core.
However, Trusted Contact goes one step further. Going beyond preventing dangerous answers, it is designed so users move toward real help.
This redefines the role of AI systems. AI cannot be the final protector in crisis situations. However, it can be an intermediary that detects danger signals and facilitates connection so users are not isolated alone.
AI Service Competitiveness Moves from 'Performance' to 'Responsible Design'
This feature also has significant meaning for the entire generative AI industry.
Until now, AI competition was primarily model performance-centered. Faster responses, longer context, better reasoning, more natural conversation, and stronger coding capabilities were core competitive factors. However, as user scale grows and the scope of conversation expands into private domains, another axis of competition is emerging.
Responsible design.
When users are in vulnerable states, how should the system respond? How to design so AI does not reinforce emotional dependency? How far to intervene when danger signals are detected? What is the method to protect life without harming privacy?
Trusted Contact is one of OpenAI's initial answers to these questions. The more human-like conversations AI has, the more human-like responsibilities AI companies are required to have.
In Korea too, this feature can trigger important discussions.
Korea has high digital service usage rates and strong non-face-to-face communication culture such as KakaoTalk, communities, and SNS. At the same time, isolation, depression, and suicide risk among youth and young adults are also very sensitive social agendas. In such an environment, AI chatbots can serve as auxiliary channels for counseling, but simultaneously social standards are needed for how to handle warning signals.
In particular, in Korea the trusted contact standard is set at age 19 or older rather than the global 18, which is also noteworthy. This appears to reflect domestic adult standards and legal and cultural context.
As AI counseling services, educational AI, and AI services for youth spread in Korea going forward, the following questions will become important: When AI detects warning signals, who to notify? Among parents, guardians, friends, schools, and institutions, which connection is appropriate? How to guarantee users' consent and right to self-determination? How to prevent stigma and privacy invasion from false positives?
Trusted Contact is a technology feature, but within it are questions of law, ethics, welfare, and education together.
Design Where AI Does Not Replace People but Returns Them to People
The essence of Trusted Contact is said by its feature name. The core is not AI but Trusted, and Contact. Connection with a trusted person.
The paradox of the AI era lies here. The more sophisticated technology becomes, the more important human-to-human connection becomes.
ChatGPT can answer more questions. Can have longer conversations. Can respond more subtly. However, the most important thing AI must do in the face of someone's crisis may not be producing perfect answers. Sometimes it is sending a signal to someone in reality.
This announcement shows the direction of AI safety. Moving from AI that blocks dangerous answers to AI that reduces dangerous isolation.
Trusted Contact is an important turning point emerging in an era when generative AI has begun dealing with human vulnerability. Not AI replacing people, but a structure that helps people reach people. That is the most important meaning of this feature.
