OpenAI publishes policies to protect teens with AI | Keryc
Today, March 24, 2026, OpenAI publishes a set of prompt-based safety policies to help developers build appropriate protections for teens. They're designed to be used with the open-weight model gpt-oss-safeguard and aim to turn safety goals into operational rules that actually work in real systems.
What was published
OpenAI released safety policies structured as ready-to-use prompts for reasoning models like gpt-oss-safeguard. What does that mean in practice? That you—developers and product teams—get clear templates to turn broad risk definitions into classifiers you can apply to user-generated content.
The first version covers concrete, high-risk areas for teens, including:
Graphic violent content
Graphic sexual content
Harmful body ideals and dangerous behaviors related to body image
Dangerous activities and viral challenges
Romantic or violent roleplay
Age-restricted goods and services
These prompts can be used for both real-time filtering and offline content analysis. They're built to fit easily into existing workflows and adapt to different use cases.
Why this matters now
Opening models with free weights democratizes innovation, but it also increases responsibility. If anyone can run powerful models, how do we make sure young people aren't exposed to specific risks? That's exactly the gap these policies aim to close: they help translate high-level safety goals into concrete, operational rules.
Many teams—even experienced ones—struggle to convert broad principles into precise rules. That leads to incomplete protections, inconsistent enforcement, or filters that are too strict and block legitimate content. These policies offer a reusable minimum floor for the whole ecosystem.
How to use it in practice
Think of an educational app with a chatbot, or a social platform with teen rooms. With these policies you can:
Implement a classifier that flags and blocks graphic content in real time.
Run periodic audits via offline analysis to spot emerging risk trends.
Integrate product-designed responses: clear warnings, referrals to help resources, parental controls.
It’s not just about pasting a prompt. The useful part is combining these policies with product design choices: user controls, teen-friendly transparency, monitoring systems, and age-tailored responses. You can also translate them, extend them to other risk areas, and adapt them to your audience's cultural context.
OpenAI worked with outside organizations like Common Sense Media and everyone.ai to broaden coverage and edge cases, so this isn’t an isolated exercise: it’s collaboration between experts and the community.
"One of the biggest gaps in AI safety for teens has been the lack of clear, operational policies that developers can build from..."
Robbie Torney, Head of AI & Digital Assessments, Common Sense Media
"Efforts like this that make youth safety policies more operational are valuable because they help translate expert knowledge into guidance that can be used in real systems."
Dr. Mathilde Cerioli, Chief Scientist at everyone.AI
Limitations and recommendations
These policies are a starting point, not a complete safety guarantee. Every product has different risks and audiences: what works for a learning app might not be enough for an open social network.
Practical recommendations:
Tune and test the policies with real data from your product.
Combine them with design controls, transparency for teens, and reporting options.
Keep continuous monitoring and metrics to detect false positives and negatives.
Participate in the community: the release is open source through the ROOST Model Community and the RMC repository on GitHub for feedback and contributions.
A practical step in a necessary direction
This isn't magic or a one-size-fits-all fix. It is, however, a concrete tool so developers don't start from zero when protecting young users. If you build products that reach teenagers, these policies give you a reusable, collaborative framework to improve safety day by day.
