OpenAI unveiled a new text‑generation model in February, but warned that the full system posed safety and security risks. Instead of publishing the complete algorithm, the nonprofit released a much smaller version and kept the training data and code private. Researchers described the model, later named GPT‑2, as capable of generating coherent, topic‑specific prose and ignited a broader safety debate.

The model was trained on text scraped from roughly eight million web pages, learning to predict the next word in a sequence. When fed a prompt, it can mimic newspaper style, novel dialogue, or political speeches, though outputs sometimes ramble or repeat phrases and occasionally produces factual errors. This ability to shift across genres marks a step beyond earlier narrow‑task generators.

Experts split on OpenAI’s embargo. Some argue the risk of mass‑produced fake articles justifies holding back the code, while others see the move as a publicity stunt that hampers academic research. Regardless, the episode forces the community to confront how powerful language models should be shared, and it has already shaped discussions about AI governance for future model releases.