OpenAI announced GPT-2, a language model that could generate convincing paragraphs of text, but said it would not release the full model because of concerns it could be misused. The decision prompted wide debate about how AI developers should publish powerful systems.
What happened
- GPT-2 had 1.5 billion parameters and was trained on text from around eight million web pages.
- OpenAI initially released only a much smaller version, citing risks such as automated disinformation, impersonation, and spam.
- The company adopted a staged release, publishing progressively larger versions during 2019.
- The full 1.5 billion parameter model was released in November 2019.
Why it mattered
GPT-2 was an early public signal that large language models could produce fluent text at scale, and it introduced the idea of staged release that later shaped wider AI safety practice. Critics argued the warnings were overstated, while others welcomed a more cautious approach to releasing powerful models.
Lessons for organisations
Assess how generative AI could be used against your organisation, for example in convincing phishing or fake content, and set clear policies on how staff may use AI tools. Keep an inventory of AI tools in use and review the risks of each before adoption.
Source: OpenAI
Part of our Top stories archive of headline-making events in information security, privacy, and AI. If you would like help applying the lessons to your organisation, contact us.
