Last month we released an API for developers to build on top of our latest technology: https://twitter.com/OpenAI/status/1271096720881901569. There’s been a lot of enthusiasm, but a great deal needs to be figured out as we scale up access. Here's how we're thinking about it. 1/8

We rolled out limited access to the API via a private beta, which is a way for us to understand the challenges of deploying open-ended generative models in the real world, but with mechanisms in place for safety. 2/8

Generative models can display both overt and diffuse harmful outputs, such as racist, sexist, or otherwise pernicious language. This is an industry-wide issue, making it easy for individual organizations to abdicate or defer responsibility. OpenAI will not. 3/8

We’ve shared API Usage Guidelines w/ users that include heuristics for safely developing applications. We do not support use cases which may cause physical or mental harm, including but not limited to harassment, intentional deception, radicalization, astroturfing, or spam. 4/8

As we gain more experience operating the API in practice we expect to expand and refine these categories. We’re actively enforcing our terms of use. 5/8

Before applications go live, they’re also subject to a review with the OpenAI team. Our technical enforcement is catching up to our policies, and the process is improving. We still have many policies to create. 6/8

We’re deploying a content filtration system as a starting point to allow a developer to control what content can be returned. Technology alone won't be the answer; we're working with users and our Academic Access partners for a more comprehensive approach. 7/8

We recognize these issues are complex; as an industry we need to work together to improve the way we tackle them. New projects that expand OpenAI’s work in these critical areas are underway, building on existing work in the field. Research & product outputs to come. 8/8