OpenAI's AI So Good at Faking News, They Won't Release It

In February 2019, OpenAI announced a new AI called GPT-2 that could generate fake news articles so convincingly that they worried it was too dangerous to share. Normally, OpenAI releases its research for anyone to use, but this time they held back, saying the risk of malicious use was too high. The AI had been trained on a massive dataset: about 10 million articles from Reddit, totaling 40 GB of text—enough to fill 35,000 copies of Moby Dick.

The decision sparked a debate. Some praised OpenAI for being responsible, while others argued that withholding the model wouldn't stop bad actors and that open research is better. The concern was that such AI could be used to spread disinformation, impersonate people, or manipulate public opinion. At the time, it was a wake-up call that AI could be used for harm, not just good.

Looking back, GPT-2 was just the beginning. Since then, AI has become even more powerful, with models like GPT-3 and ChatGPT. The same fears about fake text have only grown. Today, we see AI-generated deepfakes, phishing emails, and social media bots. The 2019 warning seems prescient—and the debate over how to balance openness and safety continues.

For everyday people, this means we need to be more skeptical of what we read online. Not everything is written by a human. AI can now mimic writing styles, making it harder to tell fact from fiction. The good news is that AI also helps detect fakes, but it's an ongoing arms race. Understanding this history helps us navigate today's AI-filled world.