OpenAI Introduces Dedicated Under-18 Safety Tests With Its Latest ChatGPT Model

OpenAI has updated ChatGPT with new versions of its models and, alongside that release, has published for the first time a set of evaluations designed specifically to measure model behaviour against safety standards for users under 18.

The company states that for users it believes may be under 18, it applies age-specific provisions that are more restrictive than those for adults. These cover sexual content, emotional reliance, eating disorders, and access to age-restricted goods and services. The model is trained with additional safety data intended to prevent romantic roleplay, discourage age-restricted challenges, and stop it from positioning itself as a substitute for real-world relationships.

According to OpenAI, when the system recognises signs that a teenager may need additional support, it is designed to reinforce healthy boundaries and encourage connection with parents, caregivers, teachers, counsellors, or other trusted people. System-level protections limit exposure to sensitive content, prompt breaks during extended use, and give parents tools to manage a child's experience.

The new under-18 evaluations draw on difficult, production-derived examples, including adversarial cases relating to self-harm, eating-disorder behaviours, and graphic violence. OpenAI notes these focus on long-tail risks and should not be read as estimates of how often such behaviours occur in ordinary use. The company also reported a statistically significant regression on a self-harm evaluation for one model, while stating it did not observe an increase in undesirable responses during online testing.

The original reporting is by OpenAI and can be read at https://deploymentsafety.openai.com/gpt-5-6-august-update.