On September 23, Anthropic said roughly 950 Claude-based agents had identified a CRISPR-like enzyme system. Days later, a University of Copenhagen biologist claimed his team had been working on the same thing for about four years, and that they used Claude in their research. The dispute is still open.
According to coverage in Smithsonian Magazine, Tribune, and The Print, the system is called ART (array-associated reverse transcriptases): a reverse transcriptase, a neighboring gene, and a long array of regularly spaced DNA repeats, an architecture reminiscent of the CRISPR arrays in bacterial defense, found mostly in bacteriophages. Anthropic reported that about 950 Claude agents searched sequence databases for around 21 hours, using roughly 210 million tokens. For the technical details, see our earlier article on the finding.
Mario Rodríguez Mestre, a biologist at the University of Copenhagen, maintains that his team discovered the system a year before the announcement, after roughly four years of unpublished work. According to the reports, the team regularly used Claude to write code and draft manuscripts, and therefore shared information from their research in those conversations. His concern is that this data may have influenced Claude’s conclusions.
Anthropic, according to The Print, denies training on user transcripts and says it has no knowledge of published data on ART. Nothing reported proves that the Copenhagen team’s data reached the system, and nothing rules it out either: this is a claim by one of the parties, not a verified finding.
That an analysis of this size took 21 hours shows the scale agents offer for pattern search. But attribution gets complicated: when a result is “discovered by AI,” the question of who contributed what, and with which data, stops being trivial. The dispute mixes two separate issues: scientific priority and the confidentiality of what you type into an assistant.
For companies: (1) review each AI provider’s terms of use and data retention policies before pasting in unpublished research, code, or designs; (2) use enterprise plans with contractual no-training guarantees if you work with intellectual property; (3) keep dated documentation of which discoveries or developments are your own, so you can demonstrate priority in a dispute; (4) set an internal policy on what information is never shared with external assistants. The practical lesson doesn’t depend on who is right in this case.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel