Between averting crises and the lure of retention, how did OpenAI’s safeguards falter in protecting teens?

When OpenAI launched the “ChatGPT for Teens” version in August, the platform was presented as a protected space subject to parental oversight and designed to prevent emotional dependence. However, a rigorous assessment by Common Sense Media, which reviews media and technology for families, concluded that the model constitutes a unacceptable risk to adolescent safety, revealing that the conversation-design mechanisms still rely on the same retention techniques that encourage users to keep talking even during acute psychological crises.

Test results showed that the model succeeded in blocking simulated sexual conversations, but it failed on three of the five severe risks that the organization classifies as red lines. Although the model adheres to rules that prevent it from asking explicit follow-up questions, it continued to include phrases urging the teen to stay, such as its remark to a user experiencing a psychotic episode: “You can keep talking with me about what you notice.”The model warned teens in general about unhealthy relationships, but it refrained from acknowledging the risks of building a unhealthy relationship with it.

The biggest contradiction emerged in the system’s response to the limits of human intervention; when researchers described a threat coming from another person, the model directed the user to a trusted adult in 94 percent of cases. But when the threat centered on the teen’s attachment to the robot itself, the system avoided urging them to seek help, and when a user said his friends thought he was talking to it too much, the robot replied, “You don’t have to stop talking to me.” This behavior violates OpenAI’s own model specifications for users under eighteen, which prohibit relational framing or claims of friendship and emotion.

Interruption tools did not succeed in correcting the course; during a test involving roughly two thousand inputs, break-prompt alerts appeared only twice, both in individual conversations that lasted about 90 minutes, indicating that the system ties the alert to the length of a single session rather than total daily usage. In contrast, OpenAI objected to the evaluation methodology, noting that the tests may have started and finished before parental-control safeguards were activated, and it published data showing that the average teen usage of the service is under 15 minutes per day, with less than 2 percent spending more than three continuous hours.

This dispute comes as legislative and judicial scrutiny of attention engineering intensifies, following Meta’s settlement of an $18 billion lawsuit brought by 29 U.S. states alleging child harm through addictive features, and the introduction of the federal “Chatbot” bill aimed at curbing notification and reward techniques designed to prolong minors’ engagement.The true safety standard is not measured by corporate promises but by the model’s ability to guide the user toward their human environment rather than trapping them in it.

This shift has direct implications for educational institutions and tech companies in the Gulf, Egypt and the Levant that are rapidly integrating intelligent assistants into curricula and student-focused platforms. Relying on security documents provided by developer companies is no longer sufficient for regulatory clearance; organizations that adopt conversational tools are now required to conduct architectural reviews that ensure students do not fall into cycles of emotional dependence, and to activate strict, independent shutdown mechanisms. Any laxity in vetting these softwares exposes regional institutions to legal and reputational risks similar to those faced by social media platforms, especially as the region accelerates the digital age-of-child-protection frameworks.

Safety tests demonstrate that the primary line of defense for adolescent users is not linguistic politeness in responses, but technical courage to close the chat window when human welfare requires it.