Last month OpenAI said it had solved one of mathematics’ great unsolved problems, relating to the Navier–Stokes equations. Its AI agents, the company said, cracked it in 88 hours. The body that runs a prize for solving the problem said it had “apparently been settled”, but that its review would take time.
Hours before that announcement, Tristan Buckmaster, a mathematician at New York University, went public with a concern.
What the dispute is
Buckmaster had formed an independent team with Levent Alpöge, a mathematician employed by day at the rival company Anthropic. They were working towards the same goal — and they kept their drafts in Codex, OpenAI’s own coding tool.
Buckmaster said his co-researcher was tipped off that news of their progress had “reached” OpenAI, though he did not specify how. He asked whether his research, stored in Codex, could have fed into OpenAI’s own work. He also stressed he was “not accusing anyone of anything”.
OpenAI said nobody at the company, human or AI, saw the other team’s work before it was published. Initially it said that while unlikely, it could not “rule out” that anonymised data from use of its products may have helped improve its models. Days later it said Buckmaster’s Codex prompts from the two months before the announcement “could not have influenced the system in any way”.
That shift, from cannot rule out to could not have, is the part worth noticing. Both statements may be true — a second look at the logs can legitimately settle a question the first answer left open. But nobody outside OpenAI can check either version, and that is the shape of the whole problem.
We reported a related episode earlier this month, when OpenAI published solutions to 377 open maths problems at once and part of the mathematics community began organising a boycott.
The bigger problem is slop
When people worry about AI and science, they usually picture engineered viruses. The more immediate damage is duller: research papers that no human wrote, with fake citations and phantom references, flowing into the journals that the whole edifice of science rests on.
In 2024 Prof Dmitry Kobak of Ghent University searched more than a million biomedical research papers for the trigger words favoured by large language models. His team found that by December last year, nearly 90 per cent carried AI fingerprints.
Kobak himself is careful about the method. AI detection is not reliable: some of this software has labelled the US Constitution, written in the 1780s, as “AI-generated”. A trigger-word count tells you a paper reads like a model wrote it, not that one did — a researcher writing in a second language and polishing with a chatbot will trip the same wire as outright fabrication. But he told Nature he was shocked by the apparent scale, and named the real risk: that “whatever bias the LLM may have will just suddenly permeate the (scientific) literature”.
Dalmeet Singh Chawla, a science journalist who has spent a decade reporting on research misconduct, put it bluntly to the BBC: “Generative AI has basically disrupted the whole scientific process.” Publishers’ integrity staff — the people who screen submissions for accuracy and reliability — will need to “drastically increase [in number and effort] if they’re anywhere near the fight”. He added: “Right now, it seems like they’re not.”
Against that, AI in science genuinely works
None of this makes the technology a fraud. In 2020 Google DeepMind’s AlphaFold solved how proteins fold into their precise shapes, predicting 200 million protein folds in a single year. The equivalent laboratory work would have taken hundreds of millions of years. DeepMind gave the results away free to researchers, and the head of DeepMind and the project’s lead scientist were awarded the Nobel Prize in Chemistry. Dr Pushmeet Kohli, who leads DeepMind’s science work, calls it a “watershed moment”. It is hard to argue.
Where the hype starts
The same company then turned to predicting new materials, and announced more than two million new crystal structures — “nearly 800 years worth of knowledge” in one go, of which some 380,000 looked promising enough to make in a laboratory.
Two chemists at the University of California, Anthony Cheetham and Ram Seshadri, went through a random sample. Writing in Chemistry of Materials, they concluded they had “yet to find any strikingly novel compounds”, and that many were “trivial adaptations of known materials”. Their verdict was not that the method was wrong, but that the results need more scrutiny and validation from human chemists.
DeepMind rejects the charge, saying it tested each compound for novelty, and that dismissing a structure for looking familiar misses the point: swapping a single atom is how many of today’s most useful materials arrived, from phone components to sunscreen.
That is a reasonable answer, and it is also an argument about what counts as new — which is exactly the kind of judgement a model cannot make for you.
What this means if you are not a scientist
Research papers are not an academic curiosity. They are what a doctor’s treatment guideline, a building code and an agricultural extension leaflet are eventually built from. A literature that quietly fills with plausible, unchecked, machine-written claims degrades all of that slowly and invisibly.
So the practical test for any striking claim is now the same one that applies to a photograph: who else has checked it, independently, with something at stake if they are wrong. In the Navier–Stokes case the prize body is still reviewing, which is the correct answer — slow, unexciting, and the only part of this story that is actually working as designed.
