When Can We Say AI Has Made a Scientific Discovery?

Summary

Anthropic recently said it had launched a molecular biology lab in which Claude agents propose ideas and human scientists conduct experiments. The company said that 950 agents identified, after 21 hours, a previously uncatalogued repeating pattern around a known enzyme. The system did not find a new DNA sequence, and the announcement did not establish what the pattern does or whether it has practical significance. Biologists argued that recognizing an unusual gene cluster is only an early step; understanding its function is where much of the scientific discovery lies. University of Copenhagen biologist Mario Rodríguez Mestre also said his team had already found the pattern, while Anthropic denied learning from his Claude conversations. The dispute highlights a broader problem: AI companies increasingly describe systems as making discoveries rather than as tools used by scientists. That framing can obscure the legitimate value of narrowing 200,000 candidates to a small set for human investigation, while forcing progress into a simplistic breakthrough-or-failure debate. The article draws a similar parallel with OpenAI’s claim that its agents solved a million-dollar mathematics problem, which critics questioned on grounds of importance and possible uncredited use of prior work, not correctness. It argues that companies should set a high evidentiary bar so that genuinely new biological mechanisms receive appropriate recognition.