GPT-2
OpenAI's 2019 language model, initially withheld as too dangerous to release, which set the template for staged AI releases.
GPT-2 was OpenAI’s language model from 2019, and by today’s standards it is tiny and not very good. It matters for what happened around it: OpenAI announced it, then declined to publish the full version, saying it was too dangerous because it could be used to mass-produce convincing fake text. The model was released in stages over the following months, and the feared wave of automated disinformation did not arrive in the form predicted.
The episode is still cited constantly, by both sides of every safety argument. Critics use it as the original example of a lab generating publicity by calling its own product dangerous. Defenders point out that staged release, external testing and capability thresholds all became normal practice afterwards. Either way, when a company says a new model may be too risky to ship, someone will bring up GPT-2 within the hour.