Release strategies and the social impacts of language models

Abstract

Large language models have a range of beneficial uses: they can assist in prose, poetry, and programming; analyze dataset biases; and more. However, their flexibility and generative capabilities also raise misuse concerns. This report discusses OpenAI’s work related to the release of its GPT-2 language model. It discusses staged release, which allows time between model releases to conduct risk and benefit analyses as model sizes increased. It also discusses ongoing partnership-based research and provides recommendations for better coordination and responsible publication in AI.