We're committed to developing AI responsibly. Safety is built into everything we do, from model training to deployment.
Safety considerations are integrated from the earliest stages of model development, not added as an afterthought.
We're open about our models' capabilities and limitations, and we publish our safety research.
We work with researchers, policymakers, and the broader community to advance AI safety.
We constantly monitor, test, and improve our safety measures as AI technology evolves.
Our safety approach is multi-layered, providing defense in depth:
Models are trained with safety in mind using RLHF (Reinforcement Learning from Human Feedback), constitutional AI principles, and carefully curated training data.
Built-in filters detect and prevent harmful outputs. The Moderation API scans content before and after generation.
Clear terms of service and acceptable use policies define appropriate use. Violations result in account restrictions.
We monitor for misuse patterns and have rapid response protocols for emerging threats.
Regular adversarial testing to identify and address vulnerabilities.
View reports →Detailed documentation of model capabilities, limitations, and safety evaluations.
View model cards →If you discover a safety issue or see concerning uses of our API, please let us know immediately. We take all reports seriously.
Report an IssueAI safety is a shared responsibility. While we work hard to build safety into our models and systems, developers play a crucial role in deploying AI responsibly.
We encourage you to: