With GPT-2, one of our key concerns was malicious use of the model (e.g., for disinformation), which is difficult to prevent once a model is open sourced. For the API, we’re able to better prevent misuse by limiting access to approved customers and use cases. We have a mandatory production review process before proposed applications can go live. In production reviews, we evaluate applications across a few axes, asking questions like: Is this a currently supported use case?, How open-ended is the application?, How risky is the application?, How do you plan to address potential misuse?, and Who are the end users of your application?.

We terminate API access for use cases that are found to cause (or are intended to cause) physical, emotional, or psychological harm to people, including but not limited to harassment, intentional deception, radicalization, astroturfing, or spam, as well as applications that have insufficient guardrails to limit misuse by end users. As we gain more experience operating the API in practice, we will continually refine the categories of use we are able to support, both to broaden the range of applications we can support, and to create finer-grained categories for those we have misuse concerns about.

One key factor we consider in approving uses of the API is the extent to which an application exhibits open-ended versus constrained behavior with regard to the underlying generative capabilities of the system. Open-ended applications of the API (i.e., ones that enable frictionless generation of large amounts of customizable text via arbitrary prompts) are especially susceptible to misuse. Constraints that can make generative use cases safer include systems design that keeps a human in the loop, end user access restrictions, post-processing of outputs, content filtration, input/output length limitations, active monitoring, and topicality limitations.

We are also continuing to conduct research into the potential misuses of models served by the API, including with third-party researchers via our academic access program. We’re starting with a very limited number of researchers at this time and already have some results from our academic partners at Middlebury Institute, University of Washington, and Allen Institute for AI. We have tens of thousands of applicants for this program already and are currently prioritizing applications focused on fairness and representation research.


版权声明:本站原创文章,由 追风者 2023-12-03发表,共计2104字。
转载说明:声明:本站内容均来自互联网,归原创作者所有,如有侵权必删除。 本站文章皆由CC-4.0协议发布。