In late September, Anthropic announced that its AI model Claude had identified a previously uncharacterized enzyme system in bacteriophage DNA, one with features reminiscent of CRISPR. A few days later, a researcher at the University of Copenhagen said he had found the same thing first, and raised an uncomfortable question about how the AI got there.

What Anthropic announced

On September 23, 2026, Anthropic unveiled a new life sciences research group and lab, along with what it called its first result. According to coverage of the announcement, around 950 Claude agents spent about 21 hours searching a large DNA sequence database for unusual reverse transcriptases, the enzymes that copy RNA into DNA. They gathered more than 200,000 of them and narrowed the field to 20 candidates for deeper study.

One agent spotted a long array of evenly spaced DNA repeats sitting next to an unusual reverse transcriptase gene. Anthropic named the system array-associated reverse transcriptase, or ART. It has three parts: the reverse transcriptase, a neighboring partner gene, and the repeat array. Early lab work reportedly showed that the array is read out as several short RNAs. What ART actually does is still unknown. Anthropic noted that the reverse transcriptase itself had been seen in earlier studies, but said Claude appeared to be the first to recognize the array and the partner protein as parts of a distinct system.

The challenge from Copenhagen

Mario Rodriguez Mestre, a researcher at the University of Copenhagen, told The New York Times that his team had worked on these enzymes for four years and had identified this specific ART about a year ago. His group calls the proteins “jumbotrons” because they were found in jumbo phages, large viruses that infect bacteria. He has not yet published the details.

His central point is not that the system was unknown before Anthropic’s announcement. It is a question about data. Mestre says he used Claude extensively during the project and shared dissertation and manuscript drafts with it. He worries that information may have been used to train future models. He also describes the similarity between the two projects as “striking.”

Mestre also has a relevant history. He worked on jumbo phages in 2022 while consulting for ReNAgade Therapeutics, and patents on reverse transcriptases filed by that company list him as a co-inventor.

Anthropic’s response

Anthropic told the Times that Claude was not trained on any user transcripts and that its molecular biology team has no access to them. The company also says its team found ART in raw data from a 2022 paper on a jumbo phage, a study Mestre was not involved in. Anthropic has not shared the exact prompts its researchers gave Claude.

What we can and cannot conclude

Mestre himself says he cannot show that the AI drew on his work. Two groups reaching the same result from the same public data is not unusual in science, especially when the underlying sequences are already published. Without the prompts, the agent logs, or Mestre’s own unpublished data, an outsider cannot tell which explanation fits best. The dispute does highlight a broader tension: researchers increasingly use AI assistants on unpublished work, and the rules about how that material is handled are not always clear to the people sharing it.

It also shows why priority claims in the age of AI agents will need more transparency, such as shared logs and clear statements about training data, to be settled convincingly.

Sources and further reading