OpenAI: AI Policies to Protect Teens

Security Policies for Adolescents in AI

OpenAI recently announced the implementation of new security policies aimed at helping developers create safer artificial intelligence (AI) experiences for adolescents. These policies are based on prompts and are designed to work with the open-weight security model, gpt-oss-safeguard. The goal is to simplify the transformation of security requirements into usable classifiers in real-world systems.

OpenAI's initiative is part of a commitment to democratize access to powerful AI models while ensuring that safety and innovation develop in tandem. Developers need not only access to high-performing models but also the necessary tools and policies to deploy them safely and responsibly. To achieve this, OpenAI has collaborated with trusted external organizations such as Common Sense Media and everyone.ai.

An Approach Tailored to Adolescents' Needs

It is recognized that adolescents and adults have different safety needs, with the former requiring additional protections. The new policies aim to help developers account for these differences to create experiences that are both empowering and appropriate for young users.

A Continued Commitment to Youth Safety

OpenAI has long been committed to developing AI that expands opportunities for young people while ensuring their safety. In this context, the company has updated its Model Spec, which defines the expected behavior of OpenAI models, to include principles specific to those under 18. Product-level protective measures, such as parental controls and age prediction, have also been introduced to better safeguard young users. OpenAI's Teen Safety Blueprint calls for industry-wide protections.

The release of the new security policies builds on these solid foundations. They are made available to developers to support them in deploying protections for adolescents and to democratize access to the open-weight ecosystem.

Clear and Usable Policies

While security classifiers like gpt-oss-safeguard can detect harmful content, they require clear definitions of what that content is. One of the main challenges for developers is to define policies that accurately capture the specific risks to adolescents and can be applied consistently. Even experienced teams may struggle to translate high-level security objectives into precise operational rules, which can lead to gaps in protection or excessive filtering. Clear and well-defined policies are essential for effective security systems.

Operationalizing Adolescent Safety

To tackle this challenge, OpenAI is releasing a set of security policies tailored to the common risks faced by adolescents. These policies are informed by a thorough review of research on adolescents' developmental differences. They are structured as prompts usable with gpt-oss-safeguard and other reasoning models, allowing developers to apply consistent safety standards.

The initial policies cover various sensitive content areas, such as graphic violent content, graphic sexual content, harmful body ideals, dangerous activities, romantic or violent role-playing, and adult-only goods and services. They can be used for real-time content filtering and for offline analysis of user-generated content.

Collaboration with External Experts

OpenAI has worked with external organizations such as Common Sense Media and everyone.ai to develop these policies. Their expertise has been crucial in defining the content scope to cover, strengthening the prompt structure, and refining the edge cases to consider during evaluation. This work reflects an ongoing effort to collaborate with experts to improve how AI systems support youth.

A Starting Point for Enhanced Safety

The policies proposed by OpenAI are designed as a starting point, not a comprehensive solution for adolescent safety. Each application presents unique risks and contexts, and developers are best positioned to understand the risks their products may pose. OpenAI encourages developers to adapt and extend these policies according to their specific needs and to combine them with other protective measures.

A defense-in-depth approach is essential for building safer AI systems. The policies draw on OpenAI's internal experience but do not reflect the entirety of the company's internal policies or protective measures.

Encouraging Collaboration and Iteration

OpenAI is releasing these policies as open source via the ROOST Model Community to encourage collaboration and iteration. Developers and organizations can adapt these policies to their specific applications, translate them into different languages, and extend them to cover other areas of risk. Over time, this should contribute to a more robust foundation for implementing security policies in AI systems.

To get started with gpt-oss-safeguard, it is available for download from Hugging Face.