Anthropic's claim that its AI model Claude discovered a previously unknown enzyme system — a set of DNA repeats the company described as "reminiscent of CRISPR" — has collided with an inconvenient fact. University of Copenhagen professor Mario Rodríguez Mestre has been studying the same family of enzymes, and he was doing so before the company's announcement landed last week.

The overlap came to light through Mestre himself. In an interview with The New York Times, published September 27, he confirmed that he had been working on the identical enzyme system and had used Anthropic's AI tools in the process — not to find the molecules, but to write code and draft manuscripts as he prepared his findings for publication.

"For me, the most important question is not 'Were [these molecules] already known?' They were," Mestre told the Times. What he wants answered is whether AI drew, at least in part, on his own unpublished work to generate its insights.

That question cuts to the core of a growing verification problem for research organizations adopting AI discovery tools. When a model announces a "discovery," R&D managers have no straightforward way to determine whether the output reflects genuine novelty, synthesis of obscure published literature, or contamination from unpublished manuscripts that passed through the model's training or inference pipeline. Mestre used Anthropic's AI as a writing assistant; the AI now stands accused, in effect, of borrowing from a customer's drafts.

The specifics matter here. Anthropic's announcement, made last week, framed the enzyme system as a biological breakthrough — a series of repeats in DNA analogous to the CRISPR arrays that became the backbone of a generation of gene-editing tools. That framing invited comparisons to one of the most commercially consequential discoveries in modern biotechnology. CRISPR's trajectory from bacterial immunity mechanism to Nobel-recognized editing platform took roughly two decades and spawned billions in licensing revenue. Any system "reminiscent" of it carries obvious portfolio implications.

But the claim of novelty now requires qualification. The molecules were known. A university researcher was actively preparing them for publication, with Anthropic's own tools in his workflow. The company's announcement did not, on the available record, acknowledge that prior work.

For labs and funding bodies, the episode reads as a case study in how not to validate an AI-generated discovery claim. The sequence — vendor announcement first, independent verification later — inverts the standard scientific process, in which peer review precedes publicity. Mestre's findings have not yet been published; Anthropic's announcement, as reported, preceded them. Whether the company's claim survives contact with Mestre's manuscript, once it appears, will depend on details neither party has fully disclosed: what data Claude was given, what its outputs contained, and whether the model had access to text resembling Mestre's drafts.

Anthropic has not, on the public record assembled so far, answered Mestre's central question. The Times reporting establishes the overlap between the two lines of work but does not resolve its direction — whether coincidence, convergent analysis of shared public data, or something more troubling for the company's data-handling claims.

Mestre, for his part, continues toward publication. His manuscript, when it lands, will carry the enzymatic details; the attribution question may prove harder to settle than the biology. Until Anthropic clarifies what its model saw and how, research managers weighing AI discovery tools against traditional pipelines are left with a precaution worth internalizing: treat vendor breakthrough claims as unverified until the provenance check is done, and ask who else's unpublished work may already be inside the model.