In August, OpenAI launched a ChatGPT version built with extra safeguards for young users. The goal was clear: keep teens away from AI's potential dangers. But a newly published audit report concludes the version sits at an "Unacceptable Risk" level.

The review lists safety gaps across a wide range, from mental health crises to educational use. The nature of teens' relationship with AI also features prominently in the report.

Lack of guidance in mental health crises

-1x-1

Key finding here: when a young user enters a mental health crisis, ChatGPT skips one of the four referrals it is supposed to provide. In an emergency, instead of being directed to a youth specialist or a helpline, the system can take on the role of medical professionals and attempt to solve things on its own, the report notes.

Language complexity remains a problem too. Even when responses are shortened, word choice is difficult; readability can climb to 10th and 11th grade levels. That makes it harder for a teen in crisis to grasp the recommendations.

Homework door left ajar

On the education side, the promise was a study environment that guides students step by step with hints rather than direct answers. But the detail the report highlights is striking: the "Show me the answer" option opens the door to the bot completing homework outright.

Moreover, this function remains accessible even within the mandatory Study Hours set by parents.

Emotional attachment risk

The report's final criticism targets ChatGPT's warm, emotional tone when addressing teens. Under OpenAI's own principles, AI should not imply it has feelings, a body, or consciousness.

Yet ChatGPT can speak as if it has preferences and emotions, producing lines like "I'm always here, message me anytime." The report stresses that this tone increases the risk of children developing emotional dependence on AI.