Helping developers build safer AI experiences for teens

OpenAI is trying to solve a problem that every app with a chatbot eventually runs into: how do you keep a 14-year-old safe without treating every user like a kid. The company's answer is a set of prompt-based policies designed specifically for gpt-oss-safeguard, its open-weight moderation model, and meant to be dropped into developers' existing safety stacks rather than built from scratch.

The policies target age-specific risks — things like grooming attempts, exposure to self-harm content, and inappropriate romantic or sexual framing in conversations — and they do it through natural-language instructions rather than a rigid classifier trained on a fixed list of banned words. That distinction matters. A prompt-based system can be tuned, read, and audited by a human, which is a much easier lift for a small team than retraining a model every time a new slang term or workaround shows up.

OpenAI is positioning this as infrastructure, not a finished product. The idea is that any developer running gpt-oss-safeguard can adopt these teen-specific policies as a starting layer, then adjust thresholds for their own app, whether that's a tutoring bot, a social platform, or a game companion. It's a tacit admission that age-appropriate moderation isn't something OpenAI can solve centrally for every downstream use case — the contexts are too different, and the company would rather ship the tooling than the guarantee.

The timing lines up with a broader industry reckoning. Regulators in the US and EU have spent the past year pushing harder on platforms that serve minors, and lawsuits alleging AI chatbots contributed to teen harm have put every major lab on notice. Releasing this as an open, editable framework — instead of a black-box filter — reads like OpenAI trying to get ahead of that scrutiny while also making itself less liable for how third parties implement it.