OpenAI Said GPT-2 Was Too Dangerous to Release. Should We Have Believed It?
A researcher's old frustration with OpenAI's 2019 safety announcement raises a question still worth asking: when an AI company warns the world about its own technology, who really benefits?
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 and Claude.
- Stanford researcher John Thickstun found the announcement unhelpful and the stated risks overblown.
- The episode reveals an incentive worth watching: AI companies that talk up danger may also be talking up power, and drawing 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 and answer questions, then announced it while refusing to share it. The model was GPT-2. Too risky for public hands, the company said.
The concern, as OpenAI framed it, was abuse: bad actors writing fake news at scale, flooding the internet with convincing misinformation. Reasonable fear, in theory.
In practice, Stanford researcher John Thickstun was left cold. Writing in The Guardian, he recalls feeling annoyed. Without access to the actual model, he couldn't study it or learn much from it. The warning was loud; the substance was 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 it's 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's not a conspiracy. It's just an incentive worth noticing.
GPT-2 was eventually released in full, in stages, later in 2019. The catastrophic misuse OpenAI feared didn't materialise in any obvious, documented wave. It became a research tool, and its successors became the engine inside products used by hundreds of millions of people today.
We covered the broader pattern of AI hype and accountability in our piece on open model ownership from 17 July, where the gap between what companies claim and what researchers can verify came up again.
Does this mean safety warnings are always wrong?
No. Some AI risks are real and some warnings deserve serious attention. The point isn't to dismiss every caution from every lab.
What matters is who is speaking, what they stand to gain, and whether outsiders can actually examine the evidence. 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, pause. Check what researchers can verify without the company's cooperation. Look at who funds the next investment round after the headlines run. Ask whether the described danger is specific or just vivid.
Curiosity and scepticism aren't opposites of taking AI seriously. They're what taking it seriously actually looks like.
