When can AI be said to have made a scientific discovery? The Anthropic-CRISPR case

Anthropic announced that earlier this year it set up a molecular biology laboratory in which Claude agents read and form hypotheses about complex biology problems, while human scientists run in the lab the experiments those agents propose. The company presented this system as responsible for its "first discovery": a set of 950 agents that, after 21 hours of work, did not find a new DNA sequence but instead flagged a repetitive pattern around an already known enzyme, a pattern that, according to Anthropic, had not been catalogued until now. In its statement, the company described that pattern as "reminiscent" of the finding that gave rise to CRISPR, the gene-editing technology that "has already transformed science and medicine", a comparison that sets expectations about the achievement very high.

That way of presenting the finding has irritated part of the biology community. A post that went viral, written by a biologist and later backed by the chair and chief executive of the pharmaceutical company Eli Lilly, argued that "finding a strange cluster of genes and repeats is usually the easy part; the hard part, where the real discoveries come from, is working out what that system actually does". On this reading, Claude's agents would have helped with routine laboratory tasks, but that does not amount to a discovery as such.

The episode grew even more complicated when Mario Rodríguez Mestre, a biologist at the University of Copenhagen, said the following weekend — as reported by the New York Times — that his team had already identified that same pattern earlier. Mestre, who regularly talked with Claude as part of his work, wondered whether the Anthropic team might have benefited from those conversations. Anthropic has denied it, but Mestre has announced he will stop using Claude in any case.

The original analysis in MIT Technology Review, by James O'Donnell in The Algorithm newsletter, pinpoints the underlying problem: AI companies do not present their systems as tools at the service of scientists — comparable to a microscope or a supercomputer — but insist that it is the AI itself that "makes" the discoveries. For many, that framing clashes with the way science actually advances, which tends to emerge from collaboration between people and from an ever-growing arsenal of tools, not from a single agent claiming the credit alone.

This way of framing results also fuels scepticism towards advances that are genuine. The article itself notes that narrowing 200,000 candidates down to a handful of options worth studying is no small achievement: it is legitimate scientific work, and it is significant that a general-purpose chatbot can do it, even if guided by humans who then ran the experiments. The problem arises when the test of success is reduced to whether or not Claude "made" a discovery on its own: that binary framing — either it is a historic milestone or it is worthless — leaves little room to acknowledge real intermediate progress.

That same pattern of dispute played out weeks earlier with OpenAI, which claimed its own agents had solved a mathematics problem worth a million dollars. Shortly afterwards, many AI critics were sharing an article asking whether that particular problem was really the one of greatest interest to the mathematics community; the piece did not dispute that the solution was correct, but its actual relevance within the field. Added to this was a mathematician's accusation that the models might have used part of his work without attribution, leaving the public with the feeling that either OpenAI had acted less than cleanly or the result was not as important as claimed — or both at once.

The biologist behind the criticism of Anthropic closed his post with a recommendation aimed at the AI companies themselves: set the bar high now, so that when an AI really does discover a fundamentally new biological mechanism, everyone is able to appreciate the magnitude of the achievement. The article, however, ends on a sceptical note: as long as those running Anthropic and OpenAI compete to outdo each other in headlines and announcements, raising the bar for what counts as a true scientific advance is probably the last thing on their minds.