Biologist questions whether AI 'borrowed' his work for major discovery; expert explains the rules

Anthropic announced that its system Claude identified an unusual family of biomolecules, namely enzymes in phages (viruses that infect bacteria), and named them array-associated reverse transcriptases (ARTs). The discovery was presented as an example of AI initiating a biological discovery.

But Mario Rodríguez Mestre, a doctoral student in computational biology at the University of Copenhagen, raised a question: did Claude independently recognize the importance of these enzymes, or was it influenced by unpublished work that had been shared with it? According to The Scientist magazine, Mestre shared his thesis and manuscript drafts with Claude. Mestre says his team members have been studying these enzymes since 2022, calling them "jumbotrons."

Anthropic denies that the model was trained on user conversation records and denies that its biology team had access to those records. Existing reporting cannot independently resolve the dispute between Mestre's questions and Anthropic's response, and there is no established evidence that his unpublished materials were used in the discovery.

Mestre is listed as one of the inventors on a patent for engineered "retrons," which also lists reverse transcriptase sequences from jumbo phages, and its U.S. application was published in July 2023. This documents earlier jumbo-phage reverse transcriptase work.

But the patent document alone cannot prove that Mestre's team had fully characterized the enzyme system described by Anthropic. This is an inference drawn from the gap between the patent's contents and the scope of Anthropic's announcement, not an established fact.

The enzyme at the core of Anthropic's announcement had previously been identified by a team of scientists at Texas A&M University in the United States. Anthropic claims to have identified a larger biological system: the enzyme, a series of repeated DNA sequences, and a possible partner protein; researchers found that these repeats produce small RNA molecules.

Some of the evidence came from reanalysis of data published in 2022 by another research team. This shows both that existing data can reveal new findings and highlights the importance of crediting the producers of the original data.

The laboratory experiments were performed by human scientists, and important questions remain unanswered. Researchers have not yet demonstrated that the enzyme is active, that it can copy the relevant RNA molecules into DNA, or what the system does in viruses.

The system's resemblance to the well-known gene-editing tool CRISPR has drawn attention, but resemblance does not mean that ARTs can edit genes; this remains an application to be studied rather than a proven one.

Anthropic's report on Claude's discovery states that in ten other search runs, the agent failed to find the repeat array. Some models could recognize the pattern when given the relevant sequence directly, but recognizing a given pattern and independently discovering it are two different things.

This limits claims about Claude's sustained autonomous discovery process. However, it does not indicate that the biological discovery itself is wrong. Anthropic's report has not yet been peer-reviewed.

Before sharing unpublished research with AI tools, scientists should consult their university's guidelines and receive training in responsible use. A paid subscription does not automatically guarantee that research remains private or is excluded from model training.

Anthropic's rules differ for its personal and commercial services: commercial services exclude inputs and outputs from training by default, but exceptions may apply when users give permission or submit feedback. Its current policy alone cannot explain what actually happened to the information Mestre says he shared.

Keeping files on one's own computer does not mean everything stays there: datasets and computations can remain on the researcher's computer, while instructions sent to the AI and its replies are processed by Anthropic. Researchers should verify what information left their computer, how long it is retained, and whether it can be used to train models. Confidential materials should be shared only when institutional rules and permissions allow.

The discovery of ARTs raises interesting biological questions, but further experiments are needed to determine what the Claude system actually did. The controversy also highlights a broader question: how do we ascribe scientific contributions when AI is involved.

In this case, credit should acknowledge the researchers who produced the original data, identified these enzymes, and revealed the connections between them. Announcements about AI discoveries should state these human contributions as clearly as they describe AI achievements.