We launched open weight fashions to democratize entry to highly effective AI and assist broad innovation. On the similar time, we consider security and innovation go hand in hand, and that builders ought to have entry to succesful fashions in addition to the instruments and insurance policies to deploy them safely and responsibly. We developed these insurance policies to assist builders of their security efforts to guard younger customers, with enter from trusted exterior organizations together with Frequent Sense Media(opens in a brand new window) and everybody.ai(opens in a brand new window).
We acknowledge that teenagers and adults have totally different wants, and that teenagers want extra protections. These insurance policies are designed to assist builders account for these variations and construct experiences which might be each empowering and applicable for youthful customers.
At this time’s launch builds on that basis. We’re making these security insurance policies obtainable to builders to assist them in deploying security protections for teenagers and serving to democratize entry throughout the open weights ecosystem.
Whereas security classifiers like gpt-oss-safeguard can detect dangerous content material, they depend upon clear definitions of what that content material is. In apply, one of many largest challenges builders face is defining insurance policies that precisely seize teen-specific dangers and may be constantly utilized in actual techniques.
Even skilled groups usually wrestle to translate high-level security objectives into exact, operational guidelines, particularly because it requires each material experience and deep AI data. This may result in gaps in safety, inconsistent enforcement, or overly broad filtering. Clear, well-scoped insurance policies are a crucial basis for efficient security techniques.
To handle this problem, we’re releasing a set of security insurance policies(opens in a brand new window), tailor-made to widespread dangers confronted by teenagers and knowledgeable by cautious overview of present analysis about teenagers’ distinctive developmental variations. These insurance policies are structured as prompts that may be instantly used with gpt-oss-safeguard(opens in a brand new window) and different reasoning fashions, enabling builders to extra simply apply constant security requirements throughout their techniques.
The preliminary launch consists of insurance policies masking:
- Graphic violent content material
- Graphic sexual content material
- Dangerous physique beliefs and behaviors
- Harmful actions and challenges
- Romantic or violent roleplay
- Age-restricted items and companies
These insurance policies can be utilized for real-time content material filtering, in addition to offline evaluation of user-generated content material.
By structuring insurance policies as prompts, builders can extra simply combine them into present workflows, adapt them to their use instances, and iterate over time.
This work displays an ongoing effort to collaborate with consultants and the broader ecosystem to enhance how AI techniques assist younger individuals.
“One of many largest gaps in AI security for teenagers has been the dearth of clear, operational insurance policies that builders can construct from. Many occasions, builders are ranging from scratch. These prompt-based insurance policies assist set a significant security flooring throughout the ecosystem, and since they’re launched as open supply, they are often tailored and improved over time. We’re inspired to see this sort of infrastructure being made obtainable broadly, and we hope it catalyzes extra shared youth-safety beginning factors throughout the trade.”
—Robbie Torney, Head of AI & Digital Assessments, Frequent Sense Media
“Efforts like this that make youth security insurance policies extra operational are invaluable as a result of they assist translate professional data into steerage that can be utilized in actual techniques. Content material insurance policies are an essential first step, and so they additionally open the door to broader work on how mannequin conduct can form youth-relevant dangers over time. Impressed by this work and our personal analysis, everybody.ai(opens in a brand new window) has additionally created an preliminary behavioral coverage centered on dangers like exclusivity and overreliance.”
—Dr. Mathilde Cerioli, Chief Scientist at everybody.AI
The insurance policies are meant as a place to begin, not as a complete or last definition or assure of minor security. Every utility has distinctive dangers, audiences and contexts, and builders are finest positioned to grasp the dangers that their merchandise and AI integrations might current. We strongly encourage builders to adapt and prolong these insurance policies based mostly on their particular wants and mix them with different safeguards corresponding to product design selections, consumer controls, teen-friendly transparency, monitoring techniques and considerate, age-appropriate responses.
We consider a layered protection in depth strategy is important to constructing safer AI techniques. These insurance policies draw from our inside expertise, however they don’t mirror the total extent of OpenAI’s inside insurance policies or safeguards.
Builders and organizations can adapt these insurance policies to their particular purposes, translate them into totally different languages, and prolong them to cowl extra threat areas. Over time, we hope this contributes to a extra sturdy and shared basis for implementing security insurance policies in AI techniques.