Anthropic’s ‘AI Scientist’ Discovery Claim Challenged: What Happened

Short answer: Anthropic’s big “AI scientist” claim is now a public dispute, not a debunking. On September 23, 2026, Anthropic announced that roughly 950 of its Claude AI agents, working for about 21 hours, had flagged a previously uncharacterized enzyme system in viruses — called array-associated reverse transcriptases, or ARTs — that the company said it discovered with only high-level direction from human scientists. Days later, a computational biologist, Mario Rodríguez Mestre of the University of Copenhagen, told The New York Times that his team had identified the same systems back in 2022 and had shared unpublished research on them with Claude. Anthropic has denied training its models on user conversations and says its biology team had no access to Mestre’s work. The result is a genuine scientific argument about provenance and originality — one that neither side has fully settled, and that has left the broader claim about AI-driven discovery sitting in uncomfortable middle ground.

This article breaks down what Anthropic actually announced, who is challenging it and on what grounds, how Anthropic responded, and — most importantly — what still has not been proven by anyone. If you are looking for a clean “AI discovered a new gene-editing tool” story or a clean “the whole thing was a fraud” story, you will not find either here. The evidence supports neither.

What Anthropic Actually Claimed

On September 23, 2026, Anthropic published a preprint — not a peer-reviewed paper — and an accompanying announcement describing what it called a Claude-led discovery. The company had set up a new life-sciences research group and a molecular biology wet lab in the San Francisco Bay Area earlier that year, and this was the lab’s first public result.

The experiment worked like this. Human researchers gave Claude agents a high-level instruction: search a large database of DNA sequences for interesting new examples of reverse transcriptase (RT), an enzyme that copies RNA into DNA. From there, according to Anthropic, the agents worked on their own. Approximately 950 Claude agents spent about 21 hours analyzing data — consuming roughly 210 million tokens — examining more than 200,000 reverse transcriptases, narrowing the pool to the most compelling candidates. One agent flagged a repeating pattern of DNA sequences sitting next to the gene for an unusual-looking RT. Anthropic’s scientists then analyzed and tested the finding in the wet lab.

The company named the system array-associated reverse transcriptases (ARTs). In brief, the ART loci consist of a repeat array, an RT gene with an unusually long N-terminal domain, and a partner protein of unknown function. The repeat-array structure has some resemblance to CRISPR, the bacterial defense system that scientists turned into a Nobel Prize-winning gene-editing technology in 2020 — and Anthropic was not shy about that comparison. CEO Dario Amodei wrote on X that the company had announced “the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism,” adding that its precise function and utility were not yet clear. In a statement accompanying the research, MIT and Broad Institute molecular biologist Feng Zhang, who was not involved in the work, called the finding “genuinely intriguing” and an encouraging example of how AI agents can contribute to biological discovery.

Three caveats were built into Anthropic’s own announcement and should not be skipped:

1. The work was a preprint. It had not been through peer review.
2. Nobody knows what ART does. Anthropic explicitly stated the system’s biological function — and whether it has any biotechnological utility at all — remains unknown.
3. The reverse transcriptase enzyme itself had been observed before. The claimed novelty was its association with the distinctive repeat array and downstream partner protein, which together define a system nobody had characterized.

Even before the dispute erupted, independent coverage was careful. Shares of gene-editing companies fell shortly after the announcement — markets moved on the CRISPR comparison — while scientists reading the work noted that an uncharacterized system with an unknown function is a long way from a new gene-editing tool. Anthropic said it does not plan to run clinical trials and does not intend to compete with pharmaceutical companies.

The Challenge: A Biologist Says He Got There First

Days after Anthropic’s announcement, Mario Rodríguez Mestre — a computational biologist who recently completed his doctoral research at the University of Copenhagen — told The New York Times that his team had studied the same enzyme systems for years and had shared unpublished analyses of them with Claude. His account, later expanded in an October 8, 2026 CNN report, forms the core of the challenge.

Mestre says he first identified the enzymes in 2022 while hunting for new reverse transcriptases in jumbo phages — oversized viruses that prey on bacteria. His team calls the systems “jumbotrons.” He briefly consulted for ReNAgade Therapeutics, which filed a 2023 patent covering several reverse transcriptases, including ones he found, and he is named as a co-inventor. His group has not yet published its work.

The critical part of Mestre’s claim is the connection to Claude. For roughly three years, he and his colleagues used Claude for coding, drafting manuscripts, and other routine research work. Mestre says he gave Claude drafts of his dissertation and a manuscript on the jumbotron systems. Records he shared with the Times, including Slack messages, appear to show the AI was deeply involved in the research. His position, as he put it to CNN: the systems in question are essentially the same finding — not merely two groups studying related protein families.

At the same time, Mestre has been careful about what he is not alleging. He told CNN directly: “I am not claiming that Anthropic deliberately took our work. I cannot demonstrate that.” His argument is narrower and, in some ways, more difficult for Anthropic to dismiss. He offered two possible explanations: either information related to his work somehow reached the model through training, or — the explanation he says he currently finds more plausible — there was considerably more prior scientific knowledge and human direction behind the search than the phrase “autonomous discovery” suggests. In other words, the finding may be real and interesting, but the claim that an AI independently discovered it may not survive scrutiny.

Several context points matter here. Mestre has no direct proof that his unpublished material influenced Anthropic’s result — he has said so himself — and he has since moved some of his work to other AI models, suggesting he takes the concern seriously regardless of the outcome. It is also worth noting that independent researchers converging on the same result is a normal, even expected, part of how science works. The unusual part is the allegation that unpublished work fed into the AI side of the pipeline, which would make the word “independent” difficult to defend.

Anthropic’s Response

Anthropic has rejected the substance of Mestre’s suggestion. According to reporting in The New York Times, the company said Claude was not trained on user transcripts and that its biology team had no access to those transcripts — meaning, in Anthropic’s account, there was no route by which Mestre’s unpublished drafts could have shaped the experiment.

The company has maintained its core claim: that the ART system was flagged by agents working from a broad prompt, and that the humans involved were limited to writing the initial instruction and running the physical lab work. To CNN‘s emailed request for comment on Mestre’s specific claims, Anthropic did not respond — at least as of the October 8 report’s publication.

That response leaves the dispute in a state familiar to anyone who follows science controversies: one side points to a mechanism by which contamination could have happened and shows records of substantial interaction with the tool; the other side denies that the mechanism exists and asserts clean provenance. Neither side has produced evidence that would end the argument. Anthropic’s architecture claims about training data are difficult for outsiders to audit independently, and Mestre’s records demonstrate that Claude saw his work without demonstrating that Anthropic’s agents benefited from it.

What Independent Scientists Are Saying

Beyond the provenance dispute, a separate current of criticism targets the significance of the finding itself — and the way it was framed. This skepticism comes from scientists with no apparent stake in either Anthropic’s claim or Mestre’s challenge.

“Interesting” is not “breakthrough”

Aaron Engelhart, an associate professor in the University of Minnesota’s Department of Genetics, Cell Biology, and Development, told CNN the results were very interesting but preliminary — they neither demonstrate gene editing nor explain the mechanism of what is actually happening. That captures the scientific mainstream’s position: the pattern is worth investigating, but nobody — including Anthropic — has shown that ART does anything useful.

Gustavo Sudre, a professor of genomic neuroimaging and artificial intelligence at King’s College London, told CNN that the announcement does show AI beginning to work with more autonomy in science — the agents received a broad direction and then exercised their own judgment about what was interesting — but he added a caution born of experience: most of the real work in science lives in the gap between spotting an interesting pattern and turning it into useful knowledge.

The “grunt work” view of AI discovery

Ilya Finkelstein, a professor in molecular biosciences at the University of Texas at Austin, argued that AI excels at “needle in a haystack” searches rather than imaginative breakthroughs — exactly what Anthropic’s experiment was. He also pointed out that the paper’s lead author is Dr. Peter Yoon, who previously worked in the lab of Jennifer Doudna, one of the architects of CRISPR and a 2020 Nobel laureate. Finkelstein’s point: Anthropic’s agents were orchestrated by people at the very top of the field, so the human contribution to the “autonomous” discovery was far from trivial. In his view, AI agents are at their best doing large-scale grunt work under expert direction.

Standards, hype, and the bigger argument

Yuval Elani, an associate professor in biochemical technologies at Imperial College London, took the skepticism further. Real biological progress, he argued, requires AI models integrated with hands-on lab work through robotics — the fiddly, one-off experimental tasks a good student performs almost instinctively — and robotics is not close to that yet. On that view, the hype around AI in biology is outpacing the actual capability.

Other researchers quoted across coverage have raised a related concern about scientific standards: treating an AI’s pattern detection plus tentative lab signals as a major discovery could lower the bar for what counts as a finding, and could shift researcher behavior toward optimizing for machine-readable novelty rather than mechanistic understanding. Several commentators have called for independent replication of the ART result and proper peer review before anyone draws firm conclusions about what the agents found.

For readers who track how AI agents are evolving more broadly, the Anthropic episode is one entry in a larger pattern — alongside disputed claims about AI solving longstanding math problems — in which bold “first” claims about machine capability immediately attract hard questions. If you want background on how AI agent systems work and what their realistic capabilities are, our explainer on AI agents walks through real examples and current limits.

Why This Dispute Matters Beyond One Enzyme

1. It tests what “independent discovery” means when humans feed models everything

Before this episode, few people outside machine-learning research had thought hard about a strange property of modern AI: millions of scientists and students use the same commercial AI tools for private work — including unpublished ideas — while AI labs run “autonomous discovery” experiments on those same tools. If unpublished work can shape a model, then “the AI discovered it independently” becomes a claim that requires auditable provenance, not just an architecture statement. Anthropic says its safeguards are sound. The dispute is whether assertions of clean provenance are good enough when billions of dollars — and scientific priority — are on the line.

2. It may change how scientists use AI tools

Finkelstein told CNN that after Mestre’s allegation, “many scientists may think twice in the future about inputting their unpublished results into these tools.” That would be a significant cost. The productivity benefit of AI assistants for researchers — coding, drafting, literature review — depends on feeding them real work-in-progress. If researchers start treating AI tools as potential contamination vectors for their priority claims, the adoption story for scientific AI gets more complicated. Readers concerned about what happens to data they upload to AI tools can read our guide on AI tool privacy settings — the principles apply to lab data as much as to business documents.

3. It highlights the competitive stakes

The ART announcement did not happen in a vacuum. Anthropic debuted Claude Science, its research tool, earlier in 2026, and its rivals have all fielded similar offerings: OpenAI has GPT-Rosalind, Microsoft has Quine, and Google DeepMind has Co-Scientist. AI executives are, as CNN put it, “betting big on biology,” presenting their models as capable of reshaping biological research the way they reshaped coding. A flagship “discovery” result is a marketing asset as well as a scientific claim — and the dispute has given every competitor a cautionary example of how quickly such assets can be contested. For a broader look at how the leading AI systems stack up, see our comparison of the major AI agents in 2026.

4. It is part of a pattern

CNN noted that Mestre’s objections echo those of mathematicians who challenged OpenAI after it announced that its models had solved a longstanding, high-profile math problem. The pattern is consistent: an AI lab announces a landmark capability result; outside experts respond with questions about prior art, human involvement, or definitions. This does not mean the lab claims are false — it means the bar for evidence has been raised, and labs will increasingly need to document provenance and human involvement with the same rigor they apply to the science itself.

What Remains Unverified — A Checklist

Because this story is developing and because both sides have said less than a full investigation would, it helps to separate what is established from what is not.

Established facts:

• Anthropic announced the ART finding on September 23, 2026, as a preprint (not peer-reviewed).
• The experiment used roughly 950 Claude agents over about 21 hours and approximately 210 million tokens.
• The system’s biological function is unknown; Anthropic says so itself.
• Mario Rodríguez Mestre of the University of Copenhagen publicly stated his team identified the same systems in 2022 and shared unpublished analyses with Claude — first to The New York Times, then to CNN on October 8, 2026.
• Mestre is named as a co-inventor on a 2023 ReNAgade Therapeutics patent covering several reverse transcriptases.
• Anthropic has denied that Claude was trained on user transcripts or that its biology team could access them.
• Gene-editing company shares fell after the announcement; Anthropic says it will not run clinical trials or compete with pharmaceutical companies.

Not yet established:

• Whether any of Mestre’s unpublished material influenced Anthropic’s experiment in any way. Mestre has no proof; Anthropic denies the mechanism exists.
• Whether the ART system is genuinely novel to science. Parts of the underlying biology were already known to parts of the field, and the preprint has not been peer-reviewed or independently replicated.
• Whether ART has any gene-editing utility whatsoever. No mechanistic picture exists.
• How much of the “discovery” was really autonomous versus shaped by expert human direction — a question of framing as much as of fact.

The honest summary: Anthropic found something interesting with powerful tools and experienced scientists, announced it as an AI-led discovery before the science was fully characterized, and is now facing a credible challenge to the “independent” part of the story that it cannot simply dismiss. None of this means the ART finding is worthless, and none of it means Mestre’s account is proven. Watch for three developments: peer review and independent replication of the preprint, any documented audit trail of training data and experiment provenance, and — if ART turns out to do something interesting in the lab — the actual biology.

A note on coverage: the sibling stories in this week’s AI research cluster cover Google’s Nano Banana 2.1 4K image model, Microsoft and NVIDIA’s hybrid-intelligence push on Windows, and the Vidu Q4 video-generation preview — each with its own set of claims worth checking against the evidence.

Frequently Asked Questions

Did Anthropic’s AI actually discover a new enzyme system?

Anthropic says its Claude agents flagged a previously uncharacterized system called ART (array-associated reverse transcriptases) in bacteriophages on September 23, 2026. But a biologist, Mario Rodríguez Mestre, says his team identified the same systems in 2022 and shared unpublished work on them with Claude. Anthropic denies any contamination. So the “discovery” part is credible but disputed — and the “independent” part is the contested piece.

What is ART, in plain English?

ART stands for array-associated reverse transcriptases. It is a DNA arrangement found mostly in bacteriophages (viruses that infect bacteria) consisting of a repeat array — a stretch of repeating DNA sequences — sitting next to an unusual reverse transcriptase gene and a partner protein. The repeat structure vaguely resembles CRISPR, which is why the announcement got attention. Nobody yet knows what ART actually does in a cell.

Who is challenging Anthropic’s claim?

Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen who recently completed his doctoral research. He says his team has studied these systems — which they call “jumbotrons” — since 2022, and that he fed drafts of his dissertation and analyses into Claude over several years. He told both The New York Times and CNN that his unpublished research describes “essentially the same finding.”

Did Anthropic steal the biologist’s research?

That is not what is being alleged — even by the challenger. Mestre explicitly told CNN: “I am not claiming that Anthropic deliberately took our work. I cannot demonstrate that.” The dispute is about whether his unpublished material could have reached the model through training data, or whether there was more prior human knowledge behind the search than “autonomous discovery” implies. Anthropic denies both possibilities.

Has the research been peer-reviewed?

No. Anthropic released the work as a preprint on September 23, 2026, and as of mid-October 2026 it had not been through peer review. Independent replication by other labs has also not been reported. Until that happens, the finding should be treated as preliminary.

Could this become a new gene-editing tool like CRISPR?

That is the speculation that moved markets, but there is no evidence for it. Anthropic itself says ART’s biological function is unknown and has not shown that it can cut, copy, or edit DNA in any useful way. Aaron Engelhart of the University of Minnesota described the results as interesting but preliminary, noting they do not demonstrate gene editing. Treat gene-editing claims about ART as unproven.

Why does this story matter if I am not a biologist?

Because it raises questions every AI user should understand: what happens to unpublished work you put into an AI chatbot, what “AI discovered X” announcements really mean, and how much human expertise hides behind claims of machine autonomy. It also matters for investors and the broader AI industry, where “scientific discovery” is becoming a key selling point for AI labs competing for funding and talent.

Sources

• CNN — “Anthropic’s announcement of a ‘Claude-led’ CRISPR-like discovery is sparking controversy. Here’s why” (October 8, 2026). cnn.com
• Smithsonian Magazine — “Anthropic Says Its A.I. Discovered a New Enzyme System That Resembles the Revolutionary Gene-Editing Tool CRISPR” (September 29, 2026). smithsonianmag.com
• Particle News — “Anthropic’s Claim That Claude ‘Discovered’ New Enzyme System Draws Challenge From Researcher.” particle.news
• AllSci — “Anthropic uses autonomous AI agents to uncover novel bacteriophage enzyme system.” allsci.com
• Digital Watch Observatory — “Anthropic says Claude discovers novel enzyme system with CRISPR-like characteristics in 21 hours.” dig.watch
• Seoul Economic Daily — “Can AI Be a Scientist? Anthropic Discovery Claim Sparks Debate” (September 29, 2026; summarizes New York Times reporting). en.sedaily.com

DroidOrbit could not independently verify the contents of Anthropic’s preprint or the records shared by Mr. Mestre. Figures and quotes attributed to outlets above come from their published reporting, accessed October 9, 2026. This article will be updated if peer review, independent replication, or new evidence changes the picture.