“Release Strategies and the Social Impacts of Language Models”, 2019-11-05 (; similar):
GPT-2 is a large-scale unsupervised language model that generates coherent paragraphs of text, first announced by OpenAI in February 2019[63]. We developed four variants of the model, ranging in size from small (124 million parameters) to large (~1.5 billion parameters).
We chose a staged release process, releasing the smallest model in February, but withholding larger models due to concerns about the potential for misuse, such as generating fake news content, impersonating others in email, or automating abusive social media content production[54]. We released the 355 million parameter model in May as part of a staged release process. We released our 774 million parameter model in August with a six-month follow up announcement, and we are now releasing our 1.5 billion parameter model.
While large language models’ flexibility and generative capabilities raise misuse concerns, they also have a range of beneficial uses—they can assist in prose, poetry, and programming; analyze dataset biases; and more. We want to release systems that will have a widely-distributed positive impact on society and have low misuse potential, and have striven to make release decisions informed by analysis, engagement, and empirical evidence.
Instead of releasing the full 1.5 billion model in February, we adopted a ‘staged release’ process. This delay of 9 months allowed time between model releases to conduct risk and benefit analyses as model sizes increased. We also hope our staged release process was helpful in allowing others time to adapt and react: giving researchers a chance to mitigate risk of potential misuse, and giving the general public time to adapt to a world in which it is prudent to mistrust everything they read a little more.
In addition to finding minimal evidence of misuse so far, several other factors contributed to our confidence in publishing our 774 million and 1.5 billion parameter models. These include what we learned about the positive social impact of beneficial uses, and what we learned through our partnerships among the AI community and through discussions across fields about establishing norms for responsible publication. This report discusses OpenAI’s work related to staged release of large models, partnership-based research, and broader issues in responsible publication that the AI community will need to address.
Overview
Staged Release
Partnerships
Engagement
Social Impacts of Large Language Models
Beneficial Use Potential
Misuse: Actor Assessment
Detecting Synthetic Text
Bias: Exploratory Research
Future Trends in Language Models
Recommendations for Publication Norms in AI
Conclusion
Acknowledgments
References
Appendices
Appendix A: Summary of Model Sharing Agreement
Appendix B: Release Timeline
Appendix C: Examples of Biases in GPT-2
Appendix D: Partner Research, Middlebury Institute of International Studies’ Center on Terrorism, Extremism, and Counterterrorism
Appendix E: Partner Research, Cornell University