OpenAI Said GPT-2 Was Too Dangerous to Release. Should We Have Believed It?

Key points

• On 14 February 2019, OpenAI announced a language model called GPT-2 but refused to release it publicly, saying the risks of misuse were too great.
• GPT-2 was the direct predecessor to the models that now power ChatGPT, Claude and similar AI chatbots.
• At least one researcher found the announcement unhelpful and the stated risks overblown.
• The episode raises a pattern worth watching: AI companies that talk up danger may also be talking up power, and attracting investors in the process.

What actually happened in 2019?

OpenAI built a language model, a type of software trained on vast amounts of text so it can write, summarise and answer questions, and then announced it to the world while refusing to share it. The company called the model GPT-2 and said it was simply too risky to put in public hands.

The concern, as OpenAI framed it, was abuse. Bad actors might use GPT-2 to write fake news at scale, or flood the internet with convincing misinformation. Reasonable fear, in theory.

In practice, at least one researcher was left cold. Writing in The Guardian, Stanford researcher John Thickstun recalls feeling annoyed at the time. Without access to the actual model, he could not study it, test it or learn much from it. The warning was loud; the substance, thin.

Who benefits when an AI company cries danger?

This is the sharper question, and Thickstun raises it directly.

Telling the world your technology is dangerous is also, quietly, telling the world your technology is powerful. Powerful technology attracts investors. Investors write large cheques. The company that just warned everyone about its scary AI is now better funded to build more of it.

That is not a conspiracy. It is just an incentive worth noticing.

GPT-2 was eventually released in full, in stages, later in 2019. The catastrophic misuse OpenAI feared did not materialise in any obvious, documented wave. The model became a research tool. Its successors, GPT-3 and then GPT-4, became the engine inside products used by hundreds of millions of people today.

Does this mean safety warnings are always wrong?

No. Some AI risks are real and some warnings deserve serious attention. The point is not to dismiss every caution from every lab.

The point is to ask who is speaking, what they stand to gain, and whether the evidence behind the warning is something outsiders can actually examine. A safety claim backed by published research and independent review is a different creature from a press release with no model attached.

When an AI company next announces that its latest system is both terrifyingly capable and too sensitive to share openly, it is worth pausing. Ask what researchers can verify without the company's cooperation. Ask who funds the next round of investment after the headlines run. Ask whether the danger described is specific and measurable, or vivid and vague.

Curiosity and scepticism are not opposites of taking AI seriously. They are what taking it seriously actually looks like.