Loading page…
Claude flagged a repeat pattern beside a known enzyme, but its function remains unknown and one researcher says his team found the pattern earlier.
Listen to this story
Anthropic’s Claude-assisted search has produced a testable candidate, not yet a demonstrated biological mechanism: researchers report that ART’s repeat array makes distinct short RNAs, but its role remains unknown. That distinction matters both scientifically and for Anthropic’s “discovery” framing.
About 950 agents searched for 21 hours, collecting over 200,000 reverse transcriptases before narrowing candidates to 3,500 and then 20 detailed reports.
Human scientists—not the agents—ran the lab experiments; short RNA production is the initial result, not proof of ART’s function.
Lucas Harrington argues that spotting a gene-and-repeat pattern is not the same as determining what the system does.
Anthropic calls a DNA pattern spotted by Claude agents a scientific discovery. Now biologists are challenging both the label and the finding’s novelty. The agents flagged repeats beside a known enzyme, and human scientists began testing the lead. But the system’s function is still unknown, while one researcher says his team had already found the pattern.
In its September 23 account, Anthropic said scientists prompted Claude to search a large DNA database for interesting reverse transcriptases, enzymes that copy RNA into DNA. Roughly 950 agents worked for 21 hours. They gathered more than 200,000 examples, selected 3,500 candidate systems and narrowed those to 20 for detailed reports.
The clue was not a new DNA sequence. The underlying enzyme had been identified before. A Claude agent noticed a long array of repeats next to its gene and examined the neighboring sequences. Anthropic calls the proposed system array-associated reverse transcriptases, or ART. It describes three parts: the enzyme, a partner gene and the repeat array. Identifying that arrangement gave scientists a target to investigate; it did not explain what the parts do together.
Human scientists performed the laboratory work. Anthropic says initial experiments found that the ART repeat array produces distinct short RNA molecules. That is a result beyond spotting a pattern in a database, but it does not establish the system’s primary function. Anthropic says further experiments are underway.
Anthropic compares the repeats’ layout to CRISPR arrays and points to other systems with similar combinations of features. Those comparisons help explain why its scientists chose to pursue ART. They do not show that ART works like CRISPR or will become a useful tool. For now, the finding is a promising experimental lead rather than an established account of a new biological mechanism.
Biologist Lucas Harrington’s objection is about what counts as discovery. Finding an unusual cluster of genes and repeats, he argued, is different from working out what the system does. The distinction does not make the agents’ search worthless: sorting a huge pool of candidates into a small set worth testing is useful scientific work. It does put pressure on Anthropic’s description of Claude as the discoverer.
A second challenge concerns priority. Mario Rodríguez Mestre, a biologist at the University of Copenhagen, says his team had already discovered the pattern, according to a New York Times account cited by MIT Technology Review. Because he had used Claude in his work, he also questioned whether Anthropic learned of it through his conversations. Anthropic denies that it did. His claim and the company’s denial leave priority and the route to the finding contested.
The dispute need not erase the work between the first prompt and the lab bench. Claude produced a short list from a large search, while people chose to test the lead and ran the experiments. The next substantive result will be whether those experiments establish what ART does. That answer would help scientists judge the biological finding on its merits, apart from the question of how much credit belongs to the agents.
Story updates
Claude’s much-publicized biology finding faces a test beyond speed: can the agents find it again? In an October 8 assessment in The Conversation, University of Manchester researcher Robson Tramontina highlighted ten additional searches that missed the defining DNA pattern. His assessment separates an intriguing biological lead from a repeatable autonomous discovery—and concerns about unpublished research from established evidence of misuse.
In its September 23 research account, Anthropic said roughly 950 Claude agents searched for 21 hours. They gathered more than 200,000 reverse transcriptases—enzymes that copy RNA into DNA—selected 3,500 candidate systems and narrowed those to 20 detailed reports. One agent noticed an orderly array of repeated DNA beside an unusual enzyme’s gene.
The enzyme itself was already known. Anthropic’s claim concerns the surrounding system: the enzyme, the repeat array and a neighboring partner gene. It calls the system array-associated reverse transcriptases, or ART. It occurs mainly in bacteriophages, viruses that infect bacteria.
Human scientists performed the laboratory work. Anthropic says early experiments showed that the repeat array produces distinct short RNA molecules. But researchers have not shown that the enzyme is active, that it copies those RNAs into DNA, or what the system does inside the virus. Its resemblance to CRISPR does not establish gene-editing ability. Tramontina also noted that the report had not been peer reviewed.
The bottleneck is then figuring out the purpose and function of these systems using deeper analysis, expert biological knowledge, and complementary targeted experiments.
Mario Rodríguez Mestre, a computational biology PhD candidate at the University of Copenhagen, says his team has investigated the enzymes since 2022. CNET reported that he told The New York Times his team had found the same systems and pointed to a patent listing him as a co-inventor.
Tramontina draws a narrower conclusion from that patent. Its US application, published in July 2023, lists reverse transcriptase sequences from jumbo phages. That documents earlier work on the enzymes, but does not by itself establish that Mestre’s team had characterized the full system Anthropic describes. Some evidence for Anthropic’s finding also came from reanalyzing another team’s data published in 2022.
The Conversation, citing The Scientist, says Mestre shared dissertation and manuscript drafts with Claude. He questioned whether unpublished material could have influenced the result. Anthropic denies that the model was trained on user transcripts or that its biology team had access to them. There is no established evidence that his unpublished material was used.
The additional searches raise a separate question about the AI method, not ownership. Tramontina writes that some models could recognize the repeat array when given the relevant sequences directly. Recognizing a supplied pattern, however, differs from independently finding it in a large search. The failed runs limit claims of consistently autonomous discovery; they do not show that the biological finding is wrong.
Scientists interviewed by CNET praised the search’s speed, while Berger warned that prolific AI research could flood science with low-impact observations. Anthropic’s stated reason for sharing early was both to demonstrate Claude’s capabilities and to let the wider community see its work. Tramontina recommends recording the model and instructions whenever AI is essential to a method, so other researchers can check and repeat the analysis.
anthropic.comLoading discussion...
Join the conversation
Where would you draw the line?
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.