Maths Dispute Fuels Fears That AI Is Eroding Trust in Science

A dispute over who solved a famous maths problem has sharpened a wider worry: that artificial intelligence, widely expected to speed up scientific discovery, may also weaken the trust on which research depends.

Last month, OpenAI said its AI agents had cracked a long-standing problem tied to the Navier-Stokes equations in 88 hours. The body that runs a prize for solving the problem said it had "apparently been settled", but added that its review would take time.

Hours before the announcement, Tristan Buckmaster, a mathematician at New York University, raised concerns publicly. He had formed an independent team with Levent Alpöge, who works for rival firm Anthropic, and the pair had been pursuing the same goal. They kept their drafts in OpenAI's coding tool, Codex. Buckmaster said Alpöge was tipped off that news of their progress had "reached" OpenAI, though he did not say how. He asked whether work stored in Codex could have fed into OpenAI's results, while stressing he was "not accusing anyone of anything".

OpenAI said nobody at the company, human or AI, saw the other team's work before publication. It first said it could not "rule out" that anonymised product data might have helped improve its models, though that was unlikely. Days later it said Buckmaster's Codex prompts from the two months before the announcement "could not have influenced the system in any way". What happened is still not fully clear.

Critics say the episode points to a quieter risk than the doomsday scenarios often associated with AI. Chatbots are known to hallucinate, inventing films or quotes, and there is growing evidence the same is happening in research. Dalmeet Singh Chawla, a science journalist who has covered research misconduct for a decade, says generative AI has "basically disrupted the whole scientific process" and that publishers' integrity teams will need to grow "drastically".

The technology has had clear successes. Google DeepMind's AlphaFold predicted 200 million protein folds in a year, work that would have taken hundreds of millions of years in laboratories. The firm gave the results to researchers for free, and its head and the project's lead scientist won the Nobel Prize for Chemistry. Other efforts have drawn scepticism. DeepMind claimed more than two million new crystal structures, but chemists Anthony Cheetham and Ram Seshadri sampled them and found no "strikingly novel compounds", calling many "trivial adaptations of known materials". DeepMind rejects the charge, saying each compound was tested for novelty.

Journals face a separate problem. Prof Dmitry Kobak of Ghent University found that by December last year nearly 90% of more than a million biomedical papers he examined carried signs of AI writing, though detection tools are imperfect. Stanford researchers found up to 17% of text in peer review reports at four major computer science conferences had been substantially modified by a language model. A Cornell team found about 147,000 phantom citations last year.

Publishers are responding. Elsevier has built around 100 signals to flag papers for human review. Nature's journals ban AI-generated images and tell referees not to upload manuscripts to AI systems, and arXiv bans researchers for a year if they submit invented references. Elsevier's Prof Sarah Main says "research-grade AI" should keep the researcher "in the driving seat".

DeepMind's Dr Pushmeet Kohli argues against abandoning AI, comparing that to refusing calculators, but says someone must take accountability: "You can't just specify the problem and let the machine write the paper and submit it as true." Prof Andrea Pauli of the Vienna BioCenter used AI responsibly to help identify a protein on zebrafish eggs that points to a lock-and-key mechanism for sperm entry. Meanwhile, Dr Oliver Vince of Basecamp Research is collecting DNA to build an AI that speaks the language of biology, saying there is "a technically feasible path to curing cancer".