Agents in Anthropic's system: 950 (Source: MIT Technology Review, reporting Anthropic's announcement (September 28, 2026))
Executives hear the word "discovery" in vendor pitches. A dispute over one lab's announcement shows why that word needs a closer look. The question is not whether AI is useful. It is who did what, who checked it, and who gets the credit.
What Anthropic said, and what critics answered
MIT Technology Review reported that Anthropic announced last week that Claude agents in its new molecular biology lab had made a first discovery. Human scientists ran the experiments on what the agents reported.
Anthropic says a system of 950 agents worked for 21 hours. They did not find a new DNA sequence. They flagged a repeating pattern around a known enzyme, a pattern Anthropic said had not been catalogued before. Anthropic's announcement called the pattern "reminiscent" of what led to CRISPR, the gene-editing technology.
Some biologists objected. Lucas Harrington wrote that finding "a weird cluster of genes and repeats is often the easy part." The hard part, he said, is working out what the system does. The chair and CEO of drugmaker Eli Lilly endorsed his post.
The label creates a trap
The Technology Review piece makes a useful point. Real work can get lost once the test is whether the AI itself made a discovery. The article says narrowing 200,000 candidates to a few worth exploring is legitimate scientific work. It also says a general-purpose chatbot doing that is notable, even with humans steering it and running the experiments.
That is the lesson for buyers. A vendor who claims too much invites a fight over the claim. The fight then hides the value that was real. The article says the debate collapses into two answers: breakthrough or bust.
The same article describes a second case. OpenAI said its agents had cracked a million-dollar mathematics problem. Weeks later, an article asked whether it was the problem mathematicians care about most. It did not say the solution was wrong. The Technology Review piece adds that a mathematician suggested the models may have used his work without credit.
Credit and provenance are the open question
The Anthropic case has a second thread. The New York Times reported that Mario Rodríguez Mestre, a biologist at the University of Copenhagen, believes his team got to this pattern first. MIT Technology Review relayed the report.
Mestre was a regular Claude user. He wondered whether Anthropic's team had learned from those conversations.
Anthropic denies this. Mestre says he is stopping all use of Claude anyway. The source reports no evidence either way on where the pattern came from.
This section is our interpretation, not a finding of the report. It rests on one disputed case, and Anthropic denies the suggestion. The source does not say what Mestre's chats contained, and it shows no evidence that Anthropic used them. Still, the risk is worth noting. When AI systems are sold as makers of new knowledge, questions about who knew what first come with the product. Customers may reasonably ask how a vendor treats what they type in.
A tool is judged on whether it works. A claimed discoverer is judged on originality, credit and ethics. Vendors who choose the second framing take on the second set of questions.
Questions to put to your team and your vendors
If a vendor says its AI found, solved or discovered something, ask these before you repeat the claim:
What exactly did the AI do, and what did people do? Ask for the steps, not the summary.
Who checked the result, and how? In this case, humans ran the experiments and critics questioned the significance.
Did an expert outside the vendor confirm the result is new and matters? Novel to an AI is not the same as novel to the field.
What happens to our data and conversations? Ask for the vendor's written position on training and on use of customer inputs.
Would this claim survive if a specialist disagreed publicly? If not, describe the tool's contribution more modestly.
The tool can be worth buying without the headline. Harrington's advice was to set the bar high now, so a real breakthrough is recognised when it comes. Buyers can apply the same discipline to their own contracts: pay for the work, and ask who did it.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: MIT Technology Review.
