OpenAI says its new model GPT-2 is too dangerous to release.

In a landmark and controversial decision, OpenAI announced it has developed a new state-of-the-art language model, GPT-2, but will not release the full version to the public. The lab cited significant concerns that the model's advanced text-generation capabilities—trained on 8 million web pages—could be misused for generating convincing fake news, automating spam and phishing campaigns, or impersonating others online. This marks one of the first times a major AI research organization has withheld a completed model due to safety and security fears, setting a new precedent in the field.

Instead of a full release, OpenAI published a much smaller 124-million-parameter version of the model, along with a technical paper detailing its capabilities. The full GPT-2 model boasts 1.5 billion parameters, making it significantly more coherent and versatile than its predecessor. The decision has sparked intense debate within the AI community, with some praising the proactive stance on ethics and others criticizing it as alarmist or a barrier to open research. The move forces a critical conversation about the responsibility of creators as generative AI becomes increasingly powerful and accessible.