For me, the most attention-grabbing AI governance discussion of 2019 concerned responsible publication norms, and it was sparked by OpenAI’s decision to delay the release of GPT-2, a language model trained to predict the next word in a text.
First announced in a blog post and paper in February, GPT-2 (a successor to GPT, or “Generative Pre-Training”) showed a remarkable ability to generate multiple paragraphs of fairly coherent writing in a wide range of styles. But what drew even more attention than GPT-2’s performance on language generation was OpenAI’s announcement that it would not be publishing the full model. The reasoning: it might be used “to generate deceptive, biased, or abusive language at scale,” and OpenAI wanted to take this occasion to prompt discussion in the machine learning (ML) community about responsible publication norms.
The post certainly succeeded at prompting discussion. Initial reactions were mixed, with many ML researchers criticizing what was perceived as a deliberate effort to create hype and attract media attention. Many also felt that OpenAI’s strategy was damaging to academic norms of openness, making it harder to replicate and verify their work. By contrast, reactions in AI policy and governance circles were largely positive, expressing appreciation for the effort to begin developing norms around publication of research that could be used in harmful ways, even if this particular work was not especially risky.
Over the course of 2019, OpenAI continued to post about GPT-2, providing updates on their conversations with other groups and their plans going forward. In a May update, OpenAI announced that it would be releasing the model in stages—publishing a “medium” version (following the “small” version with the original post), which was succeeded by a “large” version in August and an “extra-large” version in November.
During this period, multiple researchers attempted to replicate OpenAI’s work, and several succeeded in whole or in part. In one particularly interesting case, an independent researcher named Conor Leahy announced on Twitter that he had replicated the model and intended to release it publicly, in deliberate defiance of OpenAI’s release strategy. After discussions with OpenAI and other researchers, however, he changed his mind, and decided to keep his work private.
Of course, 2019 was not the year in which the ML community agreed on firm norms around responsible publishing—these questions are complex, and will require further experimentation and debate. But against a backdrop of increasingly convincing deepfake videos, ML research being turned to authoritarian purposes, and other concerning trends, the discussion kickstarted by OpenAI stands out to me as a step in the right direction.