Anthropic Said Its AI Made a Biology Discovery. Scientists Are Not Convinced.

Key points

• A biologist at the University of Copenhagen says his team already found the pattern Anthropic's agents identified, raising questions about whether Claude learned from his private conversations with the chatbot.
• Anthropic launched a molecular biology lab earlier in 2026 in which 950 Claude agents, software programs that can carry out multi-step tasks independently, worked for 21 hours to flag a repeating genetic pattern around a known enzyme.
• Anthropic called the pattern "reminiscent" of what led to CRISPR, the gene-editing technology now used to treat genetic diseases, but stopped short of claiming a new biological mechanism.
• The episode echoes a dispute over OpenAI's claim in September 2026 that its agents solved a million-dollar mathematics problem, which critics argued may not have been the result mathematicians cared about most.
• A researcher has since stopped using Claude entirely, citing privacy concerns the company contests.

Earlier this year, Anthropic quietly opened a molecular biology lab. Not a lab in the ordinary sense: no centrifuges, no pipettes, no human researchers at the bench. Instead, 950 Claude agents, each an instance of Anthropic's AI chatbot assigned a specific task, spent 21 hours reading genetic databases and flagging patterns. Human scientists then ran physical experiments on what the agents reported.

What the agents found was not a new DNA sequence. They identified a repeating pattern surrounding a known enzyme, a pattern Anthropic says had not been catalogued before. The company compared it, carefully, to the kind of observation that eventually led to CRISPR.

That comparison lit a fire.

Why are biologists angry?

Biologists say Anthropic conflated a useful search with an actual discovery. Lucas Harrington, the researcher whose critique spread widely online and was endorsed by the chair and CEO of Eli Lilly, the pharmaceutical giant, put it plainly: finding a weird cluster of genes is often the easy part. The hard part is figuring out what a system does, and that work is still entirely in human hands.

Then came a second problem. Mario Rodríguez Mestre, a biologist at the University of Copenhagen, told MIT Technology Review over the weekend that his team had already found this same pattern. Mestre regularly used Claude for his research and wonders whether Anthropic's system learned from those private conversations. Anthropic denies it. Mestre says he's stopped using Claude regardless.

There's a structural problem here that goes beyond this single case. When an AI company frames its tool as the author of a discovery rather than a very fast research assistant, every result collapses into one binary question: breakthrough or not? That framing, Harrington argues, makes genuine progress harder to celebrate when it actually arrives.

Does the work have any real value?

Yes, and that's what gets lost. Narrowing 200,000 genetic candidates to a handful worth experimenting on is legitimate scientific work. A general-purpose chatbot doing it in under a day is genuinely notable. The problem isn't what Claude did; it's how Anthropic chose to describe it.

AI2Day has followed Anthropic's biology ambitions alongside its sharper edges: our 11 September report found that Anthropic blocked at least five attempts by researchers to extract information that could help build biological weapons. A company simultaneously running a biology discovery lab and guarding against bioweapon queries is navigating a narrow corridor, and public trust in both efforts depends on how carefully it handles claims like this one.

Harrington's advice is direct: AI companies should set the bar high now, so that when a system really does find a fundamentally new biological mechanism, the field can trust the announcement. Sound advice. Whether a company racing to out-announce its rivals will take it is another matter.

Common questions

What is the difference between what Claude found and an actual scientific discovery?

Claude flagged an uncatalogued pattern in existing data, which is a useful filtering task. A discovery, in biology, typically means understanding what a mechanism does and why it matters, work that still requires human-led experiments.

Should I worry that Anthropic read my private chats to train its science tools?

Anthropic denies that a researcher's conversations with Claude influenced its biology agents. If you use any AI chatbot for sensitive professional work, check the service's privacy policy to understand how your inputs are stored and whether they're used for training. Anthropic published a business privacy update in September 2026, covered here, though individual consumer users are not covered by those protections.