Anthropic announced last week that its AI model Claude had identified a new type of enzyme, which the company called a significant advance for AI-assisted biological discovery. The system, named array-associated reverse transcriptase (ART), resembles the CRISPR gene-editing system found in bacteria.
Questions about novelty and data use
The claim quickly drew skepticism from Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen. Mestre told the New York Times that his team had already identified the same proteins, which they nicknamed "jumbotrons," in research that spanned four years, CNET reports. They had found the specific enzyme one year ago, he said. Mestre also noted that he and his colleagues had extensively used Claude as a research tool during this project, sharing dissertation and manuscript drafts with the AI model through Anthropic's Claude Science platform.
Mestre said he cannot prove that Anthropic's AI sampled his unpublished work, but he called the similarity between the two projects "striking." His concern centers on whether information from his chats was used to train future versions of Claude. Anthropic responded that the model was not trained on user transcripts and that its molecular biology team does not have such access. The company said its researchers identified the ART in raw data from a 2022 paper in which Mestre was not involved.
OpenAI faced similar accusations of borrowing from unpublished work when it claimed to have solved the Navier-Stokes Millennium Problem, a major unsolved mathematics challenge. In that case, as with the ART discovery, questions arose about whether the AI had access to unpublished research that pointed toward the solution.
This incident highlights a growing tension in AI-assisted research: scientists increasingly rely on AI tools that may implicitly train on their unpublished ideas, creating uncertainty about attribution and originality. Mestre's team has reportedly stopped using Claude for their projects and moved to other models.