Did Anthropic’s AI Agents Make a Biological Discovery? Scientists Push Back

This story originally appeared in The Algorithm, a weekly AI newsletter. To receive articles like this in your inbox, sign up here.

Last Wednesday, Anthropic announced that it had launched a molecular biology laboratory earlier this year. There, Claude’s AI agents analyze difficult biological problems, while human scientists conduct experiments based on the agents’ reports. The company said its AI-powered laboratory had made its first discovery.

What Anthropic says its AI system found

To understand Anthropic’s claim, imagine scientists searching through a library containing millions of DNA sequences collected from across the living world. A potential breakthrough might begin with the discovery of a unique sequence that encodes an interesting enzyme. Researchers would then need to determine what the enzyme does and how it might be manipulated for a useful purpose.

Anthropic said its system, using 950 agents, identified a pattern 21 hours later. The agents did not discover an entirely new sequence. Instead, they flagged a repeating pattern surrounding a known enzyme that Anthropic said had not previously been cataloged.

Anthropic described the pattern as “reminiscent” of CRISPR, the gene-editing technology that has “already transformed science and medicine.” That language made the work sound like a major scientific breakthrough. Some biologists, however, say the announcement overstates what the AI actually accomplished.

Biologists question whether this was a discovery

In a viral post, From Wang argued that “finding those strange clusters of genes and repeats is often the easy part.” The harder task, Wang said, is understanding what the system actually does—and that is where meaningful discoveries emerge. Wang’s post was subsequently endorsed by the chairman and CEO of pharmaceutical company Eli Lilly.

In other words, the AI agent may have helped scientists with a tedious research task. That does not necessarily mean it made a scientific discovery.

Even when AI identifies important patterns in biological data that are difficult for humans to recognize, the result may not qualify as a scientific breakthrough. What is new to an AI system may be routine, unremarkable, or simply not important to biologists.

Did Anthropic’s AI learn from a researcher’s conversations?

The controversy became more complicated when biologist Mario Rodríguez Mestre of the University of Copenhagen said that his team had already identified the same pattern. The New York Times reported that Mestre, who regularly spoke with Claude during his work, wondered whether Anthropic’s team had learned about the pattern from those conversations.

Anthropic denied this. Mestre said he would stop using Claude.

AI tools are not the same as autonomous scientists

Part of the problem is that AI companies do not always present their systems as tools scientists can use, such as microscopes or supercomputers. Instead, they often claim that AI systems are making discoveries themselves.

That framing can conflict with how science actually works. New knowledge typically emerges through collaboration and the combined use of an increasing number of tools.

AI narrowing 200,000 candidates down to a small group worth investigating is no small achievement. It is legitimate scientific research. The fact that a general-purpose chatbot can perform that task is noteworthy, even when a human helps operate the system and ultimately conducts the experiment.

But when the standard becomes whether Claude itself made the discovery, the discussion can collapse into a binary debate: breakthrough or bust.

The same debate is happening in mathematics

Judging AI by whether it has “made a discovery” can also lead people to move the goalposts after an apparent achievement. Earlier this month, OpenAI announced that a team of agents had solved a multimillion-dollar mathematics problem. Within weeks, however, AI skeptics were sharing a Scientific American article asking whether it was the kind of math problem mathematicians considered important.

The article did not claim that OpenAI’s solution was wrong. Instead, it argued that the specific result might not be what mathematicians care about most. A mathematician’s accusation that the model may have used some of his research without credit further complicated the debate.

As a result, the public is left with a misleading choice: either OpenAI cheated, the solution did not matter, or both.

Why the bar for AI scientific discovery matters

That is the concern raised by biologist Lucas Harrington, who criticized Anthropic’s announcement. He argued that the bar should be set high so people can recognize how significant it will be when AI genuinely discovers a fundamentally new biological mechanism.

As OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei compete to outdo one another, their public claims may not always reflect the real challenges involved in using AI to advance science.