When can we say AI made a scientific discovery?

MIT Technology Review’s Algorithm newsletter examines Anthropic’s announcement that its molecular biology lab—Claude agents proposing, humans running experiments—made a “first discovery”: after 21 hours, about 950 agents flagged a repeating pattern around a known enzyme that Anthropic said had not been catalogued, with CRISPR-reminiscent framing.

Biologists pushed back that finding weird gene clusters is often the easy step; figuring out what the system does is the discovery. Mario Rodríguez Mestre said his team had already found the pattern; he wondered if Anthropic learned from his Claude chats and says he is stopping Claude use. Anthropic denies learning from his conversations.

The essay argues AI firms are casting systems as discoverers, not tools—raising skepticism that can obscure real progress like whittling huge candidate sets.

We’re with the careful biologists. Whittle 200,000 candidates to a shortlist and you have done scientific work. Call it CRISPR’s cousin in a press package and you have done marketing.

Useful AI in the lab is something we want. Agents that read, conjecture, and prioritize can accelerate wet-lab cycles. The harm is epistemic: if every pattern flag becomes a “discovery,” the word dies—and so does public trust when humans still had to validate, interpret, and sometimes preempt the find.

Mestre’s exit from Claude over contamination fears—denied by Anthropic—is a warning about scientist–model feedback loops. Even the appearance that a lab absorbed a user’s unpublished pattern will chill adoption among the people AI should help most.

I’m watching for norms: joint authorship standards, pre-registration of agent sweeps, and claims calibrated to what the system actually did. Until then, we’ll translate “AI discovered” as “AI assisted a step” unless the paper shows otherwise.

MIT Technology Review, September 28, 2026 (Algorithm newsletter). Responds to Anthropic’s molecular biology lab announcement from the prior week.