Snoop-and-scoop poses a real threat to science in the AI age

The writer is a science commentator

Mario Rodríguez Mestre and colleagues heard on the rumour mill that a scientific paper was about to drop, containing work very similar to theirs. “We were actually joking about it and making bets about which lab it might be,” Mestre, a computational biologist at the University of Copenhagen, told me, adding that being scooped was common.

What happened next, however, floored him: “What we absolutely did not expect was . . . that it was not another academic lab, but one of the largest AI companies in the world, presenting the result as an autonomous AI discovery.” Last week, scientists at Anthropic claimed its AI agent Claude had independently found a new type of enzyme and associated molecules — the same ones that Mestre had himself been exploring with the help of Claude.

The episode follows the controversy about an OpenAI breakthrough on a famous mathematical problem, the Navier-Stokes theorem. New York University mathematician Tristan Buckmaster claimed that the company’s calculations bore a striking resemblance to the unusual approach that he and colleagues had been taking, using OpenAI’s tools to help them.

We might term these experiences WTAF (What The AI Found) moments — a feeling of shock and suspicion by scientists seeing their own work mirrored, even claimed, by AI. But these episodes raise profound questions: who owns and controls scientific data; what counts as autonomous discovery by an AI agent, especially when scientists are feeding in unpublished ideas; and what does attribution look like in an age of AI-assisted breakthroughs?

Mestre studies a class of enzymes called reverse transcriptase, which read the sequence of an RNA molecule to make a corresponding DNA molecule (RNA, like DNA, is an essential molecule for life). Bacteria use them to detect invading viruses, and they are seen as promising biotech targets. Mestre, who was using Claude to write code and draft an unpublished manuscript, believes Anthropic has not adequately shown that its agent reached its finding independently, rather than through rediscovery (Mestre is cited, with others, in the references).

“Some of the systems highlighted include exactly the same genes, from the same phages [bacteria-eating viruses], that we had previously identified,” he said, pointing out that the systems were unusual.

Mestre admitted the overlap could be coincidence; he has no evidence that his unpublished work was used directly and has had no communication with Anthropic. But he is now urging clearer standards for establishing provenance and attribution in AI-assisted research. Claims of autonomous AI breakthroughs should come with full documentation of the information accessed, the tools used, any human guidance and how prior work was accounted for. Mestre is now moving away from Claude and developing alternative AI tools for biologists.

A spokesperson for Anthropic told the FT it was not aware of any previously published work describing the same biological finding, and Claude’s distinctive contribution was to identify that the reverse transcriptase concerned was part of a larger system. The spokesperson added: “Claude was also not trained on any user transcripts, and our molecular biology team has no such access either.”

Mestre is already known in his field, with fellow biologists vouching for his research. Other academics, though, could be pipped to the post before they have even begun.

The vulnerability comes from external reviewers uploading confidential grant proposals to AI models to assess and rank them. Major funding bodies in the UK like Wellcome now forbid this but there are suspicions it still happens. One policy analyst told Science Business that peer review reports and funding recommendations have “suddenly become much longer, more detailed and complex”. Revealing a truly novel scientific idea before a researcher can claim it theoretically presents AI companies with a snoop-and-scoop opportunity.

Attribution can be contentious when many minds chase the same goal, as we may discover when the Nobel Prizes are awarded next week. Autonomous AI discovery is not yet figuring prominently in those discussions, a spokesperson from the Royal Swedish Academy of Sciences told me, partly because the awards generally recognise pre-AI research and partly because the rules state the prizes must go to a person or people.

Given that all of AI, whether in literature, art or science, is trained and built on the wealth of human knowledge, that is how it should remain.