A University of Copenhagen scientist says Anthropic’s AI system identified biological features his team had already documented and shared through Claude, raising questions about how the company determined that its agents had independently made the discovery.
By yourNEWS Media Newsroom
A computational biologist is challenging Anthropic’s assertion that artificial intelligence agents independently uncovered a previously unknown biological system with potential applications in gene editing, saying he and other researchers had been investigating the same enzymes for years and had shared unpublished details of their work through Anthropic’s Claude models.
Mario Rodríguez Mestre of the University of Copenhagen told The New York Times that he first identified the enzymes in 2022 while searching for reverse transcriptases in jumbo phages, unusually large viruses that infect bacteria.
Anthropic announced last week that AI agents working in its biology research environment had identified what the company calls array-associated reverse transcriptases, or ARTs. The company described the work as an instance in which its AI took “the first steps of a biological discovery on its own.”
Mestre disputes the novelty of that characterization.
“So, for me, the most important question is not ‘Were ARTs already known?’ They were,” Mestre said.
After initially encountering the enzymes four years ago, Mestre joined the University of Copenhagen and continued studying them with colleagues. His group informally calls the enzymes “jumbotrons,” reflecting their association with jumbo phages.
Their research has not yet been formally published.
Mestre also previously worked briefly as a consultant for ReNAgade Therapeutics. The biotechnology company filed a patent in 2023 covering several reverse transcriptases, including enzymes Mestre had identified. He is listed as a co-inventor.
The dispute is complicated by Mestre’s extensive use of Anthropic’s technology during the same period.
For roughly three years, Mestre and his colleagues regularly used Claude to assist with coding, manuscript preparation and other research-related work. According to the materials reviewed by the Times, Mestre supplied Claude with information connected to his enzyme research, including a draft of his dissertation, a manuscript describing jumbotrons and other unpublished details submitted while carrying out routine tasks.
The newspaper said it reviewed Slack messages, scientific figures and other records documenting both the team’s work and Mestre’s interactions with Claude over several years.
Anthropic has said one of Claude’s significant contributions was identifying RNA molecules and a protein associated with reverse transcriptase inside the biological system.
Mestre said his research group had reached that conclusion more than a year earlier.
The disagreement extends beyond the enzymes themselves to an analysis of previously collected experimental data.
As part of Anthropic’s research, its AI agents were directed to explore databases containing billions of genetic sequences obtained from environments including oceans and soil. Much of that genetic material comes from microbes and viruses whose biological functions remain unknown.
Anthropic has not disclosed the exact prompts used in the experiment.
Researchers at the company supplied the AI system with five examples of reverse transcriptases. Those enzymes convert RNA sequences into corresponding DNA and can play different roles in microbial and viral systems. Bacteria can use reverse transcriptases in mechanisms associated with detecting viruses, while viruses may use them in competition with other viruses.
The AI agents then created their own software for searching and evaluating candidate sequences, according to Anthropic. The system identified potential reverse transcriptase genes inside jumbo phages that were positioned near arrays of RNA genes.
Anthropic researchers also reported that the AI independently located raw experimental data generated in 2022 involving RNA produced after a microbe was infected by a jumbo phage. The AI concluded from the data that the infected microbes produced high levels of RNA molecules associated with ARTs, a finding Anthropic said supported its hypothesis about the function of the enzymes.
Mestre said his group had already examined the same 2022 experimental data approximately a year earlier.
That analysis helped his team recognize RNA genes associated with the enzymes they call jumbotrons, he said. Mestre further said the researchers had supplied their unpublished analysis to Claude before Anthropic announced its results.
That sequence of events led Mestre to question whether information supplied by researchers could have indirectly influenced later AI systems, although he said he does not have evidence establishing that this occurred.
“My concern is that this information was used to train future versions of the models,” Mestre said. “I think it’s important to raise this possibility.”
He pointed to previous disputes involving artists and writers who have accused AI developers of training models on copyrighted or privately supplied material without authorization.
Anthropic denied that Mestre’s conversations or unpublished findings played any role in its research.
“We are not aware of any previously published work describing the ART system we recently found. Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either,” Anthropic said in a statement Sunday.
The company’s statement draws a distinction between unpublished research that may have existed elsewhere and the question of whether its own scientists or models had access to that work before making the ART findings.
Mestre’s account has received support from at least one researcher familiar with the field.
Seth Shipman, a bioengineer affiliated with the Gladstone Institutes and the University of California, San Francisco, said he knew of Mestre’s research and that enzyme systems of this type were already familiar to numerous specialists.
Shipman said jumbotrons are uncommon and that identifying their associated RNA genes can be especially difficult because the sequences may be difficult to distinguish from random stretches of genetic material.
He also said his own research group relies extensively on Anthropic tools and shares scientific work with the company.
The debate highlights a growing challenge for laboratories adopting generative AI: researchers increasingly use the same systems for drafting, programming and analysis that AI companies are simultaneously developing into autonomous scientific research tools.
A related dispute arose this month involving OpenAI after the company announced that its technology had solved a longstanding mathematical problem. New York University mathematician Tristan Buckmaster said he and a colleague had been using OpenAI models while pursuing their own solution to the same problem. OpenAI denied training its models on Buckmaster’s conversations.
Anthropic’s biological research claims arrive as the company is drawing greater attention to AI’s potential role in scientific discovery.
Chief Executive Dario Amodei discussed those ambitions Wednesday in an address to the United Nations Security Council.
“We believe this type of application will accelerate rapidly in the years ahead, and that we can make profound progress in the near future on diseases that have plagued humanity for millennia,” Amodei said.
Anthropic is also expected to pursue a public offering that could value the company at as much as $2 trillion.
Mestre, meanwhile, had expanded his use of Anthropic products shortly before the disagreement emerged. He began using Claude Science, a research-focused Anthropic product, last month.
Following the company’s announcement about the ART system, Mestre said he began winding down projects involving Claude and moving his work to other AI models.
The dispute does not resolve whether Anthropic’s agents independently reproduced findings that had already been made elsewhere or whether unpublished information supplied through Claude could have influenced later systems. Anthropic maintains that its models were not trained on user transcripts and that its molecular biology team had no access to Mestre’s private research, while Mestre maintains that the central biological findings attributed to the AI had already been developed by his group.
The original report details the documents and communications reviewed in examining the competing accounts, while additional coverage of the dispute describes how Mestre’s research developed before Anthropic’s announcement.
Comments