OpenAI publishes AI policies for teenagers

OpenAI has published safety policies designed to help developers protect teenagers when they use artificial intelligence systems. They are presented as ready-to-use instructions for gpt-oss-safeguard, an open-weight safety model capable of detecting potentially harmful content.

The proposal aims to address a less visible challenge than building a model: defining precisely what should be considered dangerous for a teenager and how a system should respond to that content.

Rules ready to integrate

The policies are written as prompts, meaning instructions a developer can enter into a model to turn safety criteria into a classifier. A classifier analyzes content and determines whether it fits into a risk category.

The first version covers six areas:

• Graphic violence
• Explicit sexual content
• Harmful body-related ideals and behaviors
• Dangerous activities and challenges
• Romantic or violent roleplay
• Age-restricted goods and services

These rules can be used to review content in real time, for example during a conversation, or to analyze user-created posts and messages afterward.

OpenAI says the format allows teams to adapt the policies to each product, translate them into other languages, and expand them with new categories. They can also be used with other reasoning models, not only gpt-oss-safeguard.

Why teenagers need their own rules

The company argues that teenagers and adults do not have the same needs when using AI. As a result, protections designed for adult users may fall short in situations involving sexuality, body image, dangerous challenges, or emotional relationships with a chatbot.

Until now, many teams had to turn broad goals, such as “protect minors,” into specific rules for their systems. That work requires knowledge of child safety, product design, and AI models. If the rules are too narrow, they may miss risks. If they are too broad, they may block legitimate conversations.

The new policies aim to provide a common starting point. OpenAI developed them with input from Common Sense Media and everyone.ai, organizations that helped define the included risks, edge cases, and structure of the instructions.

What changes for people using these tools

The announcement does not mean that every application using open models will automatically adopt these protections. Developers will have to integrate, test, and adjust them for their product’s audience and context.

OpenAI stresses that these policies are not a complete safety guarantee. An educational application, a social network, and an entertainment chatbot may face different risks. That is why the company recommends combining them with other measures:

• Controls for users and families
• Design appropriate for teenagers
• Clear information about how AI works
• Monitoring and detection systems
• Age-appropriate responses
• Ongoing review of system behavior

The company also notes that a content filter cannot solve problems related to emotional dependency, privacy, or the way a product influences its users on its own.

An open standard to get started

OpenAI has released these policies as open source through ROOST Model Community, with the aim of having other teams review, adapt, and improve them. The goal is for developers not to have to start from scratch each time they build an AI feature for minors.

The next step will be testing how they work outside controlled examples: what content they detect, what errors they make, and how they respond to different languages, cultures, and contexts. The important idea is not that a perfect filter exists, but that more products start with explicit rules and can make their safety decisions visible.