U.S. artificial intelligence company Anthropic has claimed that its AI model Claude made an independent discovery in biology. But a scientist who has studied the field has raised the possibility that his own unpublished research, shared in conversations with the AI, influenced the finding, touching off a dispute.
According to The New York Times on the 27th, Mario Rodriguez Mestre, a researcher in the field, raised the objection after Anthropic announced Claude's scientific discovery.
Anthropic recently said it had used Claude to discover a new enzyme system with properties similar to CRISPR, the gene-editing technology. Named the array-associated reverse transcriptase (ART), the system is built around reverse transcriptase (RT), an enzyme that copies RNA into DNA.
Anthropic presented the finding as a case showing that AI can generate new scientific hypotheses and insights on its own, going beyond organizing existing knowledge or carrying out instructions from human researchers.
Mestre said he had studied the enzyme for several years and had previously shared unpublished research results related to what Anthropic described as a discovery during conversations with Claude.
At the heart of the dispute is whether Claude independently arrived at a scientific fact that was not previously known, or whether information the researcher had earlier fed into the AI shaped the result in some form.
The Times said that if Mestre's account is accurate, it could raise questions about how far the term "independent discovery" by AI can be accepted, because if unpublished information provided by a researcher was used directly or indirectly in the AI's reasoning, it would be difficult to define the result as a fully independent discovery.
Anthropic rejected the claim. The company said Claude had not used the content of its conversations with users as training data for the model that carried out the research, and that the Anthropic molecular biology team behind the study also had no access to Mestre's conversations with Claude.
With the information currently available, it is difficult to establish whether the AI reached the same conclusion independently of the human researcher's earlier work. Demonstrating an AI's capacity for scientific discovery requires the ability to verify not only the novelty of the result but also what data and information it accessed and what reasoning process it followed.
The Times noted that the controversy also illustrates a new question of who owns a discovery, as generative AI comes into full use in scientific research. Until now, AI has been used as a tool to assist scientists, analyzing vast numbers of papers, predicting protein structures and searching for drug candidates. But AI companies stress that their models are advancing toward a stage where they formulate hypotheses and propose experimental directions on their own, making the question of whether credit for a discovery belongs to humans or to AI increasingly important.
In particular, if more scientists feed unpublished data or ideas into generative AI during their research, disputes over research confidentiality, intellectual property and academic credit could intensify.
The Times said no evidence has so far emerged that Claude was actually influenced by Mestre's research, while the claim that Claude reached the scientific insight entirely independently is also being contested.
