Was Anthropic’s Biology Breakthrough Borrowed From Another Scientist’s Work?

AI company Anthropic announced last week that its flagship model Claude had identified a new type of enzyme. But a researcher at the University of Copenhagen said that he found the protein first in research four years in the making.

The scientist, Mario Rodríguez Mestre, has not yet published the details of his work, but he said he used Claude extensively in his research process. In an interview with The New York Times, Mestre has now questioned exactly what information Claude used to arrive at its “discovery.”

A Protein Discovery?

The disputed enzyme is linked to a biologically ubiquitous family of proteins called reverse transcriptases (RTs), which convert RNA into DNA. In last week’s announcement, Anthropic said its model found a tandem repeat array in an RT, which it called an array-associated RT (ART), after a vast search of 20,000 candidate proteins.¹ Anthropic said the previously unidentified ART resembled the enzyme system adapted from bacteria that gives CRISPR gene editing its ability to rewrite genetic material. The only issue, Mestre said, is that researchers had already identified the system.

“So, for me, the most important question is not ‘Were ARTs already known?’ They were,” Mestre told the Times.

Instead, he is concerned that Anthropic’s AI agent utilized some of the data he had fed into chats during his research process. “My concern is that this information was used to train future versions of the models,” he added. Mestre said he and his colleagues had been working on the enzymes for four years and had identified the specific ART enzyme one year ago. His team had called the proteins “jumbotrons” because they found the RTs in viruses called jumbo phages, which infect bacteria.

On Sunday, Anthropic addressed the concerns around the discovery, writing in a statement to the Times that “Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either.”

A “Striking” Similarity

Mestre said he first worked on jumbo phages in 2022 while consulting for ReNAgade Therapeutics. Patents on RTs from that time, filed by the company, list Mestre as a co-inventor.

After leaving ReNAgade Therapeutics, Mestre continued to study jumbotrons at the University of Copenhagen. He said that he had used Claude as a research tool throughout this project, sharing dissertation and manuscript drafts with the AI model. The Times reported that documents Mestre shared with its reporters showed he had involved Claude extensively in his work. Anthropic has long targeted life science applications of its AI and announced the release of the AI tool suite Claude Science earlier this year.

In its announcement last week, Anthropic said its research team had identified the ART in raw data from a 2022 paper in which Mestre was not involved.² The company did not share any of the exact prompts its research team had fed into Claude to send it on its search. Given this lack of transparency, Mestre couldn't say definitively that the AI had sampled his work, but he called the similarity between the two projects “striking.”

This controversy is just the latest to strike AI’s forays into research, following a mathematician’s claim that AI company OpenAI had borrowed from his work to crack a major unsolved math problem.