Anthropic said its artificial intelligence had helped uncover a new biological system, but a University of Copenhagen scientist now says his team had been working on the same enzymes for years and shared findings with the company's models — a claim that complicates the question of whether the A.I. truly made an independent discovery.
What researchers say
Mario Rodríguez Mestre, a computational biologist, told reporters that he and colleagues had studied the enzyme system Anthropic calls ARTs for about four years and had regularly used Anthropic’s A.I. models to assist with coding, manuscript drafts and other research tasks over the last three years. The Copenhagen team has not yet published its work.
“So, for me, the most important question is not ‘Were ARTs already known?’ They were,” Dr. Mestre said. “My concern is that this information was used to train future versions of the models.”
Anthropic responded, in a public statement, that it was not aware of any previously published description of the ART system it reported and that its model Claude was not trained on user transcripts. The company added that its molecular biology team did not have access to user transcripts. Anthropic further said Claude’s principal contribution was identifying a set of RNA molecules and a protein linked with reverse transcriptase within the system it reported.
Why the dispute matters
The disagreement touches on several high-stakes issues for modern scientific practice: attribution of discovery, the provenance of training data for large language models, and the transparency of private-sector laboratories using A.I. to do life sciences research.
- Attribution: If academic researchers discovered or characterised the enzymes first, norms around credit and authorship may be at stake.
- Model training: Whether user-provided content is incorporated into later model versions affects both privacy and intellectual property concerns.
- Transparency: Corporate assertions about how models were trained and what data they used are critical for public trust and scientific reproducibility.
Similar controversies have surfaced recently in other fields. The New York Times and other outlets reported disputes after private A.I. developers claimed breakthroughs in mathematics that some outside researchers said mirrored their own, unpublished work.
Timeline and claims
Based on public statements from the parties involved, the basic sequence is:
| Event | Approximate timing |
|---|---|
| University of Copenhagen team studies the enzymes | About four years (ongoing) |
| Team uses Anthropic’s models in research work | Regularly over the past three years |
| Anthropic announces discovery of ARTs | Last week (company statement) |
The Copenhagen researchers say they made the same identification of RNA molecules and a reverse transcriptase–associated protein more than a year before Anthropic’s announcement, though they have not yet published those results in a peer-reviewed venue. Anthropic maintains it found the system using its own internal work and that it has not been trained on user transcripts.
Questions for industry and academia
Observers say the episode underscores a need for clearer rules about how interactions between external users and proprietary A.I. systems are handled. Key areas of interest include:
- Whether corporate models ingest or otherwise retain user-supplied research material that could later influence model outputs;
- Standards for disclosure when private firms publish findings that may overlap with ongoing academic work;
- Mechanisms for resolving disputes when unpublished but substantive research appears to predate corporate claims.
Anthropic’s assurance that it does not use user transcripts to train models will be scrutinised by researchers and regulators alike, who want both reproducibility of scientific claims and protection for researchers who collaborate with or use corporate tools as part of their workflow.
The episode is likely to prompt close attention from scientific journals, funding bodies and ethics boards that set community norms around data sharing, model training and attribution when private A.I. tools participate in life-sciences research.
The competing accounts are unresolved in the public record: Anthropic stands by its announcement, and the Copenhagen team insists it had prior, overlapping findings and is concerned about how those may have informed subsequent model behaviour. The situation illustrates the frictions that arise as A.I. systems become more deeply embedded in laboratory research.
