Anthropic's AI lab claims its first biology discovery, and biologists push back

Anthropic launched a molecular biology lab earlier this year. According to MIT Technology Review's AI newsletter, The Algorithm, the announcement came on 28 September. Claude agents read about hard biology problems and speculate about them, while human scientists run experiments on what the agents report. The company said the lab had made its first discovery.

What the system actually found matters. MIT Technology Review reports the sequence was not new. After 21 hours, Anthropic's 950 agents flagged a repeating pattern around a known enzyme, one the company said had not been catalogued before.

What the agents did, in plain terms

MIT Technology Review frames the problem as a library problem. There are millions of DNA sequences, and the pile grows as scientists sequence more of the living world. One route to a breakthrough is spotting a peculiar sequence that encodes an interesting enzyme. Then you have to work out what that enzyme does. The newsletter notes that this second step is where the difficulty lies, and where human expertise remains essential. Anthropic's system, the company says, was built to help with the first step, not the second. Yet the announcement blurred the two, according to critics.

Anthropic's announcement leaned on a comparison. It called the pattern "reminiscent" of what led to CRISPR, the gene-editing technology that, in the company's phrasing, "has already transformed science and medicine." Read that way, an army of agents had found something notable.

Biologists did not read it that way. A viral post from Lucas Harrington, a biologist, argued that "finding a weird cluster of genes and repeats is often the easy part," and that the hard part, where real discoveries come from, is figuring out what the system does. The post was subsequently endorsed by the chair and CEO of the drugmaker Eli Lilly, according to MIT Technology Review.

A prior claim, and a question about Claude

Then came a second problem. Over the weekend, Mario Rodríguez Mestre, a biologist at the University of Copenhagen, said his team had already discovered this same pattern, the New York Times reported. Mestre, who regularly chatted with Claude in his work, wondered whether Anthropic's team had learned from his conversations. Anthropic denies this. Mestre says he is stopping all use of Claude anyway.

This is the part of the story that will not resolve quickly. A company says its model found something. A researcher says he found it first, four years ago, and had discussed related work with the model. The dispute turns on training data and disclosure, not on biology, and there is no obvious mechanism for settling it in public.

The deeper argument is about framing. MIT Technology Review notes that AI companies are not presenting their systems simply as tools, like microscopes or supercomputers, but insisting the systems make discoveries themselves. To some biologists, that is incompatible with how science works, where new knowledge tends to emerge from collaboration and a growing arsenal of tools.

Why the goalposts move

There is a cost to the framing. MIT Technology Review argues that once the standard becomes whether Claude itself made a discovery, every result gets sorted into one of two buckets: breakthrough or bust. The middle ground disappears from view, even though the newsletter calls that legitimate scientific work. That is the ground where an AI narrows 200,000 candidates down to a few worth exploring.

The same pattern showed up earlier this month. OpenAI said its own agents had cracked a million-dollar mathematics problem, and a couple of weeks later nearly every AI skeptic was sharing an article asking whether it was the math problem that really mattered. MIT Technology Review is explicit that the piece did not argue OpenAI's solution was wrong. It argued the particular result may not be the one mathematicians care most about. Add an accusation from a mathematician that the models used some of his work without credit, and the public is left with two readings: OpenAI cheated, or the answer did not matter. Or both.

Harrington's closing suggestion, quoted by MIT Technology Review, was that AI companies should "set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is." The newsletter adds a blunt assessment: as Sam Altman and Dario Amodei race to one-up each other, raising the bar may be the last thing on their minds.

The rest of the week in discovery claims

Anthropic's lab was not the only item in the 29 September edition of The Download. MIT Technology Review also flagged that SpaceX's Starship reached orbit for the first time and deployed Starlink satellites, though the flight was cut short after an engine failure. It noted OpenAI scrapped a model, GPT-6.1 Astra, over safety concerns, and that Florida asked a court to block OpenAI from developing new models. A separate item said Meta's AI agent invited a stranger to a user's home while negotiating a Marketplace sale.

On the science side, the newsletter pointed to a Nature finding that old hearts appear to get younger after transplantation, and to Quanta reporting that mathematicians cracked a 55-year-old problem using randomness, completing a proof proposed in 1971.

The 29 September edition also carried a quote from the ShinyHunters hacking group, which told the New York Times it targeted the FBI to retaliate against an advisory warning about its activities: "We may sound like children whose feelings are hurt, sure, but in our game, reputation is all that matters."

The bar for a discovery

Two things have merged in the Anthropic episode, and they are worth separating. One is capability: whether a general-purpose chatbot can sift a vast sequence space and surface candidates a human would struggle to see. MIT Technology Review concedes this is notable, even with humans steering and running the experiments. The other is credit: whether the model, rather than the pipeline around it, is the discoverer.

The second question has no clean answer, which is why the argument keeps recurring. A laboratory result is a chain: hypothesis, search, candidate, assay, replication, interpretation. Anthropic's agents occupied one link. The company's announcement described the whole chain as the system's discovery. Biologists described the same event as help with grunt work.

Both descriptions can be true at once. That is precisely the problem for anyone trying to judge progress. If a company claims a discovery and a researcher says the pattern was known, the disagreement is not about whether the software ran. It is about what counts.

MIT Technology Review's own framing of the stakes is narrower than the headlines suggest. The newsletter does not say AI cannot contribute to biology. It says the industry may be setting the wrong bar, and that grand claims make real progress harder to recognize when it arrives.

Anthropic has not backed away from the announcement. The company denies that Mestre's conversations with Claude fed into its result. Mestre has stopped using the model. Neither position has been independently verified in the material available.

What remains is a test case that other labs will now watch. If Anthropic's lab produces a second result, one where the mechanism, not just the pattern, is established and reproduced by an outside group, the framing debate becomes easier. Until then, the 28 September account remains an argument about vocabulary as much as about biology.

What to watch next

MIT Technology Review will publish its 2026 list of Climate Tech Companies to Watch on 6 October, a package covering energy storage, nuclear power and transportation. That is a separate thread from the AI discovery dispute, but it shares the same editorial problem: how to judge claims of progress in fields where the timelines are long and the metrics contested.

For the AI side, the near-term signal is not another announcement. It is whether an independent laboratory reproduces a mechanism that Anthropic's system surfaced, and whether the company describes that work as the system's discovery or as the lab's.

Until that happens, the safest reading of the 28 September story is the one MIT Technology Review offers: what is novel for an AI may be routine, unsurprising, or simply not that consequential to a biologist.