OpenAI withholds GPT-2, declaring its own text generator too dangerous to release in full

OpenAI built a language model it declared too dangerous to publish, then published part of it. GPT-2 was trained to predict the next word across 40 gigabytes of internet text, producing output that the company said "adapts to the style and content of the conditioning text." For every benign application, OpenAI counted several abusive ones: fake news, impersonation, automated spam and abuse on social media. Its supporting evidence was a prompt asking the model to argue that recycling is bad. GPT-2 complied with a multi-paragraph case that recycling causes obesity, heart disease and cancer, and is "a major contributor to global warming."

The company shipped a smaller version of the model and said it would revisit the decision in six months, citing a charter clause anticipating that "safety and security concerns will reduce our traditional publishing in the future." Critics accused OpenAI of "closing off" its research and doing the "opposite of open." Others called the move a new bar for ethics. Policy director Jack Clark described "a very tough balancing act for us." Elon Musk, an early funder, was pulled into the thread and confirmed he had not been involved with the company "for over a year."