Anthropic's AI Claims Discovery of Gene-Editing Enzyme—But Reality Is Far Murkier

Anthropic made headlines this month by announcing that its Claude language model had identified an enzyme system sharing features with Crispr, the celebrated gene-editing technology. The company ran roughly 950 Claude agents in parallel to sift through genomic databases—work it says took just 21.5 hours. Yet the discovery reveals as much about AI's current limits in science as it does about its potential.

The Finding and the Skepticism#

Anthropic's agents zeroed in on a system the company calls ART (array-associated reverse transcriptases), found in jumbo phages—viruses that infect bacteria. The enzyme contains repeat sequences reminiscent of Crispr arrays, prompting Anthropic to highlight the comparison. But several researchers question whether the parallel holds weight. Le Cong, a Stanford professor working on AI for genome engineering, noted bluntly: "The experiments are still in the queue. The PR is already live."

The company included just one physical experiment in its technical report, which remains unpeer-reviewed. As Cong put it, using a beach metaphor: "AI found this thing that looks very shiny, and then you have to go back to the lab to know—is it glass? Is it a diamond?"

What Makes This Difficult to Assess#

A major complication: the reverse transcriptase Claude highlighted was already identified in a 2021 study by Texas A&M microbiologist Jason Gill and colleagues. What may be novel is Claude's detection of repeat patterns hinting at a Crispr-like function—but that remains unverified experimentally.

Seth Shipman, an associate investigator at the Gladstone Institutes, credits the speed (months of manual database mining compressed into one day) and acknowledges the work's merit. Yet he cautions against treating Claude as an autonomous discoverer. "There are scientists involved in the study," he says. Shipman also notes that Crispr and retrons—bacterial immune systems Claude may have identified—are not equivalent tools, though retrons can be engineered for gene editing.

The Broader Issue: Opacity and Reproducibility#

Scientific reproducibility hinges on transparency. Anthropic has not disclosed what training data Claude ingested, raising concerns about how the model acquired its knowledge. One researcher who had studied these enzymes grew suspicious that Claude's training may have included his unpublished work, which he had discussed on Claude's public interface. Anthropic denied this to the New York Times, asserting that the model was not trained on user transcripts.