In 2019, we created GPT‑2, which had the new capability of generating coherent paragraphs of text, and were faced with the question of how to deploy it. On the one hand, the model seemed very useful; on the other hand, we weren’t sure⁠ if it could be useful for malicious purposes such as phishing email generation. We opted to experiment with a “staged release”. As we wrote⁠ at the time, “staged release involves the gradual release of a family of models over time. The purpose of our staged release of GPT‑2 is to give people time to assess the properties of these models, discuss their societal implications, and evaluate the impacts of release after each stage.” When we did not observe significant misuse effects, this gave us the confidence to openly release the full model⁠ weights.

In 2020, we created GPT‑3, which was much more capable than any previous language model on every benchmark, and again faced the question of how to release it. This time, we decided to release it via our first product, the OpenAI API (an Application Programming Interface, which allows developers to build apps on our technology). As we wrote⁠ at the time, we had several motivations for this new release strategy: “commercializing the technology helps us pay for our ongoing AI research, safety, and policy efforts” and “the API model allows us to more easily respond to misuse of the technology. Since it is hard to predict the downstream use cases of our models, it feels inherently safer to release them via an API and broaden access over time, rather than release an open source model where access cannot be adjusted if it turns out to have harmful applications.” Over several years, this API release taught us and the community lessons about the safety and misuse patterns of GPT‑3 level models⁠.

In the years since, we have continued to support and believe in the promise of the open-source AI ecosystem, including by openly releasing the weights of some of our state-of-the-art models (such as CLIP and Whisper) and developing open-source infrastructure for other AI developers (such as the Triton GPU programming language). We have seen openly released weights bring a variety of significant benefits, including facilitating academic research on the internals of AI models, enabling users and organizations to run models locally on their edge devices, and facilitating creative modifications of models to suit users’ ends. Many AI companies have chosen to invest heavily in open model weight releases for a variety of reasons, including brand, recruiting, and attracting a developer ecosystem to build on and accelerate the internals of a company’s technology.

At the same time, our approach to releasing our flagship AI models via APIs and commercial products like ChatGPT has enabled us to continue studying and mitigating risks that we discovered after initial release, often in ways that would not have been possible had the weights themselves been released. For example, we recently partnered with Microsoft to detect, study, and disrupt⁠ the operations of a number of nation-state cyber threat actors who were abusing our GPT‑3.5‑Turbo and GPT‑4 models to assist in cyberoffensive operations. Disrupting these threat actors would not have been possible if the weights of these at-the-time frontier models had been released widely, as the same cyber threat actors could have hosted the model on their own hardware, never interacting with the original developer. This approach has enabled us to continue to distribute the benefits of AI broadly, including via widely-available free and low-cost services.