OpenAI is rolling out ChatGPT for Teens, an automatic experience for users who identify as 13 to 17, or whom its age-prediction system believes are under 18.

The teen version places tighter limits around self-harm, eating disorders, violence, dangerous activities, sexual material and romantic roleplay. It also warns before sensitive images are shared and is designed to discourage teens from treating the chatbot as a substitute for real relationships.

The most consequential change is not any single safeguard. It is that OpenAI is combining years of youth-safety work into a default experience and using age prediction to decide who receives it.

Safety no longer depends on a parent opting in

OpenAI’s age system can consider a user’s stated age, how old the account is, when and how it is used, and the general topics discussed. Adults incorrectly classified as minors can verify their age through Persona.

That automatic routing matters because the core protections apply even when no parent account is connected. But parental oversight remains limited and voluntary.

A parent and teen must exchange and accept an invitation to link their accounts. Teens can later unlink them, though the parent is notified. Parents cannot read prompts or chat histories.

They may receive alerts in a narrow set of high-risk situations, including acute self-harm concerns and certain eating-disorder concerns. These are not general notifications whenever a teen enters a harmful prompt.

Nor does the product impose a universal time limit. Parents can set optional Quiet Hours and Study Hours, while teens receive more frequent reminders to take breaks.

Homework becomes a safety feature

ChatGPT for Teens also changes how the product handles schoolwork. When a request appears to seek a ready-made homework answer, a “responsible homework reminder” redirects the conversation into Study Mode, which works through the problem step by step.

Optional Study Hours give families another way to steer use toward learning. The broader idea is that a product marketed as a homework helper should encourage understanding rather than quietly doing the assignment.

These learning tools are among the genuinely new elements. Much of the safety system predates this launch: OpenAI introduced parental controls and acute-distress alerts in September 2025, published teen-specific model rules in December, began rolling out age prediction in January and had already added parent-enabled Study Mode and stronger break reminders by July.

The launch is best understood as consolidation with automatic enforcement, not an entirely new safety architecture.

The need is clearer than the proof

Teen chatbot use is already widespread. In a weighted March 2026 US survey of 1,204 children, 71% of 13- to 15-year-olds and 77% of 16- to 17-year-olds said they had used AI chatbots.

Among the 823 chatbot users surveyed, roughly one in six had encountered material they considered inappropriate for them. After the latest such incident, 53% told nobody and only 33% told a trusted adult.

That gap helps explain the value of automatic safeguards. It also exposes the limits of parental alerts: they reach only linked families and only when OpenAI’s systems recognize one of a small number of severe situations.

Whether the new experience reliably handles those situations is still uncertain. OpenAI’s August GPT-5.6 system card includes its first dedicated evaluations for users under 18, but reports no statistically significant overall improvement over its earlier GPT-5.5 models across the evaluated categories.

One adversarial, multi-turn self-harm test produced a worse result for GPT-5.6 Sol: 0.901, down from 0.967 for GPT-5.5 Instant. OpenAI called the decline statistically significant and said it was investigating, while noting that live experiments had not shown a corresponding increase in undesirable responses.

The company also cautions that its internal benchmarks do not estimate real-world incident rates. No independent post-launch audit of the teen experience has yet established how the safeguards perform in practice.

ChatGPT for Teens therefore addresses an important product-design problem: young users often encounter AI privately, without adults knowing when something has gone wrong. Automatic defaults can narrow that exposure. But the central test is now outside the launch announcement, whether the system consistently recognizes risk during the messy, extended conversations where safety matters most.