In central London, in a converted watch works, a row of white freezers hums quietly. Racked inside are dozens of small boxes filled with barcoded glass tubes, which together hold some hundred billion genes. Dr Oliver Vince, a co-founder of the company that operates the facility, Basecamp Research, argues that just two of those boxes could carry ten times as much information as every word of text online. His ambition is to collect DNA on a vast scale and train an AI to read the language of biology, echoing the way American and Chinese technology firms built chatbots by scraping the written web.
Vince speaks for a large and optimistic camp. Plenty of people in industry and academia expect AI to supercharge research. A dispute among mathematicians shows how fast that optimism can curdle into suspicion.
A claim, and a competing draft
OpenAI recently said its AI agents had solved a celebrated open problem tied to the Navier-Stokes equations, which govern fluid flow, in 88 hours. The organisation that administers a prize for the problem said the result appeared to settle it, but warned that checking would take time. Fair enough: proofs at this level can take years to vet.
Then came the complication. On the day of OpenAI's announcement, only hours beforehand, Tristan Buckmaster of New York University raised a concern publicly. He had been working toward the same result in a partnership with Levent Alpöge, who is employed by OpenAI's rival Anthropic, and the pair had been keeping their working drafts in Codex, OpenAI's coding tool. According to the BBC's account, Buckmaster said Alpöge had been warned that OpenAI had learned how far their work had got, though he did not name the source. He asked whether material stored in Codex could have fed into the company's own effort, and was careful to say he was accusing nobody of anything.
OpenAI's answer has shifted in emphasis. It said that no one at the company, human or machine, saw the other team's work before it was published. At first it added that, while unlikely, it couldn't rule out anonymised data from use of its products having helped improve its models. Days later it went further, saying that the prompts Buckmaster had entered into Codex during the two months before the announcement could have had no influence on its system whatever.
What happened, then? Nobody outside the companies can say for certain, which is rather the issue. A claim of priority in mathematics has always rested on a paper trail anyone can read. Keep your drafts in a tool built by a competitor and it is like leaving your diary at a rival's house: the trail now runs through a company's servers and a company's assurances.
The flood at the journals
The Navier-Stokes row is one thread of a wider anxiety, that AI in science could erode the trust on which the whole enterprise depends. The more mundane version is already here. Chatbots hallucinate, inventing films that were never made and quotes nobody uttered, and there is growing evidence the same thing is turning up in research writing, in citations and claims that look plausible and are false.
Journals are meant to be the filter. The science journalist Dalmeet Singh Chawla, who has spent ten years reporting on research misconduct, argues that generative AI has thrown the process into disarray. In his view, the integrity teams who screen submissions would need to grow sharply in both headcount and effort to keep pace. Asked whether they are doing so, he is blunt: "Right now, it seems like they're not."
I'd add a caution of my own. Peer review was never built to catch fabrication at industrial volume. It assumes an author is, broadly, a person who did the work. Strip that assumption away and the system strains in places nobody budgeted for.
The case for the machines
It would be unfair to tell only the dark half. On paper, AI is exactly the right instrument for modern research. Telescopes and particle accelerators produce billions of data points, far more than any human can read, and software that picks the signal out of the noise is useful by any measure.
The showpiece is Google DeepMind, the London-based lab. In 2020 it cracked the protein-folding problem, which asks how the chains of amino acids in living things curl into the precise shapes that decide what they do. Proteins build cells, fight infection and form tissues, so shape is function. Before AI, working out a structure meant slow laboratory experiments. In a single year DeepMind's AlphaFold predicted the structures of 200 million proteins; reproducing that catalogue by conventional laboratory methods is generally reckoned to be a task of geological, not human, timescales. The company gave the results away to researchers, and both the head of DeepMind and the project's lead scientist went on to share a Nobel Prize in Chemistry. Dr Pushmeet Kohli, who leads DeepMind's science work, has called it a "watershed moment", a phrase press offices adore, though a Nobel does lend this one some weight.
What made AlphaFold trusted, though, was that its predictions could be tested against experiment and the results were public. Anyone could check. The same is not yet true of an AI-produced proof that few humans can read, or of a model trained on a proprietary pile of genomes in a London freezer.
What to watch
The immediate test is the prize body's review of OpenAI's Navier-Stokes claim. If the proof holds, the argument over who saw what, and when, will linger as a footnote or as a precedent. If it doesn't, the episode becomes a cautionary tale about announcing first and verifying later.
Either way, researchers are likely to ask harder questions before they park unpublished work in a rival's tools, and journals will face pressure to say plainly how they screen for machine-made error. Those are unglamorous fixes. They may matter more than any single breakthrough, because a discovery nobody can trust is only a rumour with good formatting.