When can we say AI has made a scientific discovery?

By James O'Donnell (The Algorithm, MIT Technology Review) · 28.09.2026.

Last Wednesday Anthropic announced that earlier this year it had launched a molecular biology lab in which Claude agents read and formulate conjectures about hard biology problems, while human scientists run the experiments those agents propose. According to the company, that AI-powered lab achieved its first discovery.

To understand what Anthropic says its system did, the author suggests imagining walking through a library of millions of DNA sequences, amassed as scientists sequence ever more of the living world. One step toward a scientific breakthrough might be finding a peculiar sequence that encodes an interesting enzyme; then you would have to work out what that enzyme does and, eventually, how to manipulate it for something useful.

What Anthropic claims its 950-agent system found, after 21 hours of work, was not a new sequence but a repeated pattern around an already known enzyme, a specific pattern that, according to the company, had not been catalogued before. Yet Anthropic's own announcement describes this pattern as 'reminiscent' of what led to the CRISPR gene-editing technology, which 'has already transformed science and medicine', implying that the finding is more significant than it appears.

These claims have angered biologists. A post that went viral, written by one of them and later backed by the president and CEO of the pharmaceutical company Eli Lilly, argued that 'finding a weird cluster of genes and repeats is usually the easy part. The hard part, and where the real discoveries come from, is figuring out what that system actually does'. In other words: the agents helped with some of the more mechanical lab work, but that does not amount to a discovery.

The case serves as a reminder that, even when AI does something impressive — such as spotting a pattern in a mass of biological data that would be hard to perceive with human eyes alone — the result itself may not represent an advance for science. What is novel for AI can be routine, unsurprising or simply of little relevance to a biologist.

To complicate matters further, Mario Rodríguez Mestre, a biologist at the University of Copenhagen, said over the weekend — as reported by the New York Times — that his team had already discovered that same pattern earlier. Mestre, who regularly conversed with Claude in his work, wondered whether Anthropic's team had learned from those conversations. Anthropic denies it, but Mestre says he will stop using Claude either way.

Part of the problem, according to the author, is that AI companies do not present their systems simply as tools scientists can use — like a microscope or a supercomputer — but insist that it is the AI systems themselves that make the discoveries. For some, that framing is incompatible with the way science actually works, where new knowledge usually emerges from collaboration and from an ever-growing arsenal of tools.

This is also generating more skepticism toward genuine progress when it does occur. Narrowing 200,000 candidates down to a few worth exploring is no small thing; it is legitimate scientific work. That a general-purpose chatbot can do that work is remarkable, even if humans helped direct it and ultimately ran the experiments. But once the standard becomes whether or not Claude 'made' a discovery on its own, everything turns into ammunition for a binary debate: breakthrough or failure.

Once you start judging AI by whether or not it has made a discovery, it also becomes tempting to move the goalposts even when it appears to score a real win. Earlier this month, OpenAI claimed that its own team of agents had solved a math problem worth a million dollars. But a couple of weeks later, nearly every AI skeptic in the author's feed was sharing an article questioning whether that was really the math problem the mathematical community cared most about. The article did not argue that OpenAI's solution was wrong, but that this particular result might not be the one mathematicians were most concerned with. Added to that was one mathematician's accusation that the models might have used part of his work without credit, leaving people with the sense that either OpenAI cheated or the solution was not that important. Or both.

That is part of what worries Lucas Harrington, the biologist who wrote the post criticizing Anthropic's announcement. He ended with a suggestion: AI companies, he said, should 'set the bar high now, so that when an AI truly discovers a fundamentally new biological mechanism, everyone appreciates the magnitude of that achievement'. But while Sam Altman (OpenAI) and Dario Amodei (Anthropic) compete to outdo each other, raising the bar for scientific discoveries attributed to AI may be the last thing on their minds.