OpenAI Withholds GPT-2 Over Safety Concerns
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI unveiled GPT-2, a language model trained on 8 million webpages that can produce coherent prose adapting to a prompt's style and content, outperforming prior text generators in versatility and contextual word-sense disambiguation.
- OpenAI refused to release the full algorithm, training data, or code, publishing only a "much smaller version" and warning that the full model could generate fake news articles, impersonate people online, and automate large-scale spam and abuse.
- Some machine learning researchers accused OpenAI of exaggerating the risks for media attention and depriving academics who lack compute resources of the ability to study the model firsthand.
- Robert Frederking, principal systems scientist at Carnegie Mellon's Language Technologies Institute, argued that embargoing results is largely futile since well-resourced actors could replicate the approach using off-the-shelf tools and rented cloud servers.
- MIT's David Bau defended the move as a "gesture" designed to spark public debate about AI ethics, saying OpenAI deserves credit for spotlighting the issue even if withholding one model won't change the long-term trajectory.
- The model's actual capabilities are more modest than alarmist headlines suggested: GPT-2 still rambles, repeats itself, botches topic transitions, and has been caught referencing impossible phenomena like "fires happening under water."
- Dominant media coverage — from Metro U.K. ("kept locked up for the good of humanity"), CNET ("so good it's scary"), and the Guardian ("brace for the robot apocalypse") — framed the decision as evidence of an imminent AI threat, a framing the article argues overstated both the model's power and the novelty of the underlying technique.
Why it matters: OpenAI's refusal to release a fully trained model marks one of the first high-profile cases of an AI lab publicly withholding research over safety concerns, potentially setting a precedent for how the industry handles dual-use algorithms. The episode also exposed a fault line in the machine learning community between researchers who see embargoes as responsible risk management and those who view them as performative gestures that primarily benefit well-funded labs over academics.
Ask SkimNews
