Anthropic’s Enzyme Discovery Claim Raises a Provenance Question

Anthropic’s claim of an AI-assisted enzyme discovery has a complication: a computational biologist says his team had studied the same enzymes for years and shared key findings with the company while using its models. The account raises a question of provenance, or where the information behind a result came from, but does not establish that Anthropic’s AI relied on the team’s work.

What Anthropic said it found

When Anthropic unveiled findings from its biology lab, its scientists said they had used AI agents to discover new enzymes with promise for biotechnology. The company calls the enzymes and associated molecules ARTs. The available account does not describe their properties or demonstrate specific applications, so the biotechnology promise remains a claim rather than a validated use case. The discovery announcement and its characterization of ARTs

That distinction matters: identifying a candidate is not the same as showing what it can do. The central dispute here is narrower still. It concerns whether the AI arrived at the result independently, not whether the supplied reporting establishes the enzymes’ practical value.

A researcher says his team already knew the ARTs

Mario Rodríguez Mestre, a computational biologist, said he and his colleagues had studied the ART enzymes and associated molecules for four years. Their findings had not yet been published at the time of the report. Mestre also said the team had used Anthropic’s models for three years to write code, draft manuscripts and handle other tasks, and had shared key findings about the enzymes with the company. Mestre’s account of the research and model use

That history creates a real attribution question, but it does not answer it. The reported overlap and sharing make it reasonable to ask what information was available to the agents and how their work was directed. The available account does not detail the agents’ prompts, data access or workflow, so it cannot show whether they independently reasoned their way to the results or were guided by prior work.

Mestre also raised a separate concern: that information from his unpublished research might be used to train future versions of Anthropic’s models. That is a concern he expressed, not evidence that the information entered training data. His concern about future model training

The claim needs a clearer account of contribution

For AI-assisted research, “discovery” can compress several different contributions: prior human findings, information shared with a model, the model’s analysis and subsequent scientific validation. The reported facts leave those contributions unresolved. They support neither a confident claim of autonomous discovery nor a conclusion that Anthropic’s models simply reproduced Mestre’s work.

The useful next distinction is between a result an AI system helped produce and one it reached independently. In this case, the source account establishes that researchers shared relevant findings while using Anthropic models; it does not establish how that information affected the lab’s result. That gap is central to judging the discovery claim, and to understanding what the claim credits the AI with doing.