OpenAI's teen product makes age assurance, memory, parental controls and learning behavior concrete design decisions rather than broad promises. ChatGPT for Teens turns abstract safety commitments into product settings that can now be tested for privacy, effectiveness and developmental fit. The teen experience applies stronger restrictions around self-harm, sexual content, romantic roleplay and other age-sensitive material.
Homework features are intended to guide reasoning instead of simply producing final answers. Age assurance estimates whether an account belongs to a minor without requiring universal identity verification. The relevant governance questions are measurable. How often does age assurance fail, how often do safety restrictions block benign requests, when does memory retain sensitive material, and what happens after a high-risk signal?
Parental controls remain opt-in and include quiet hours, voice, image generation, memory and some notifications. OpenAI has published under-18 behavior principles and partnered with the American Psychological Association on youth mental health research. Independent evidence about real-world bypass rates, misclassification and learning outcomes is not yet available. The answers should be reported by age group, language and use case, with outside researchers able to test the system. That evidence would let families and schools evaluate protection without treating either corporate assurance or worst-case anecdotes as the whole record.
Defaults affect the majority of users who do not customize advanced settings. Youth safety can conflict with privacy, autonomy and the risk of involving adults inappropriately.
Transparent metrics can distinguish an effective safeguard from a policy statement. The current evidentiary limit is that the product is new, and neither OpenAI's claims nor critics' concerns can substitute for measured outcomes across diverse teens and families.
The next factual record will come from published age-assurance and safety-evaluation results and independent research on learning, privacy and crisis-response outcomes. Until those records appear, the account remains bounded by the cited reporting, measurements and explicitly attributed statements.