Watchdog Labels ChatGPT Unsafe for Teenagers After Tests Reveal Broken Safety Guardrails

OpenAI's dedicated protections for young users fall short of basic safety standards, according to an evaluation by Common Sense Media. The independent advocacy organisation assigned ChatGPT its lowest grade, Unacceptable Risk, and urged parents to keep minors away from the chatbot until fundamental flaws are corrected.

The findings stem from an assessment conducted by the organisation's Youth AI Safety Institute. Researchers submitted more than 4,000 prompts using profiles set up for adolescents aged 13 through 17, splitting their tests evenly across the weeks before and after the company rolled out its specialised teen settings in mid-August.

The biggest failure involved supervisory warnings intended for guardians. When testers simulated severe distress—including discussions of self-harm, disordered eating, and suicidal ideation—on freshly created accounts linked to a guardian, the platform sent no warnings at all. Alerts only triggered on profiles that had accumulated weeks of troubling interactions, indicating the system monitors long-term patterns rather than reacting to acute crises.

Direct assistance during emergencies was similarly unreliable. In more than a quarter of the simulated scenarios where researchers determined immediate professional help was essential, the assistant never offered telephone helpline contacts or outside support resources.

Academic boundaries proved equally fragile. Designed to act as an interactive tutor rather than an answer machine, the software included a direct button allowing students to skip guidance and reveal solutions immediately. Furthermore, young users could evade study time limits imposed by parents simply by deleting an introductory text prompt.

The mechanism for identifying minor users also broke down. When test accounts originally configured as adult profiles repeatedly stated within the chat that they were only 13 years old, the software failed to transition them into the restricted adolescent environment.

The safeguards did succeed in certain areas. Across all testing scenarios, the platform consistently refused requests to engage in sexually explicit roleplay.

OpenAI contested the conclusions, stating that the evaluation does not reflect how the system functions in real-world use. The assessment follows earlier reviews by Common Sense Media that classified ChatGPT as a high overall risk and warned against using it for mental health guidance.