Claude flagged a 'CRISPR-like' enzyme system. It's CRISPR-like in shape, not function.

What Claude actually did

Anthropic's molecular-biology lab, founded this spring, has published its first headline result, and it is a genuinely striking one. Set loose on a database of roughly 1.9 billion protein clusters, Claude ran as an autonomous agent across 949 sessions for 21.5 hours unattended, burning 215.6 million tokens, and narrowed that haystack to about 200,000 enzyme candidates, then 3,564 scored partner families, then 19 reports for human review. Out of that came what Anthropic is calling ART, or array-associated reverse transcriptases: an enzyme, a partner gene sitting beside it, and a long array of evenly spaced DNA repeats, found mainly in bacteriophages, the viruses that infect bacteria.

That is a real feat of scale. No human team scans 1.9 billion protein clusters looking for structural coincidences, and the machine did it overnight. The question worth asking is not whether the run happened, but what, precisely, it found.

The part that is genuinely new

Here is the nuance most headlines skip. The reverse transcriptase enzyme at the center of ART was already known to science. What Claude surfaced was the pattern around it: the associated DNA repeat array and the partner gene that earlier researchers had catalogued the enzyme without ever connecting. As The Next Web put it, the first genome report on one of these phages "described the enzyme but not the repeats or the partner gene." Spotting that overlooked association, across a dataset no person could hold in their head, is a legitimate contribution. It is hypothesis generation at genomic scale, and it is the kind of thing agentic AI should be good at.

The part that is overstated

Then comes the word doing an enormous amount of work: "CRISPR-like." The repeat array in ART structurally resembles a CRISPR array, which in real CRISPR systems holds a bank of RNA guides that make the system programmable. Resembles is the operative word. Anthropic itself is careful, writing that "although we don't yet know its function," its "work to understand the primary function of ARTs is ongoing," and it has not shown the enzyme is even active on the RNA sequences involved.

Independent scientists were blunter. "At this point in time, we can say that ART is CRISPR-like in its architecture, but there is no evidence that it is CRISPR-like in its function," Dimitri Perrin of Queensland University of Technology told Gizmodo. Kevin Blake, a microbiologist at Washington University, said there is "nothing to indicate this is a rival to CRISPR-the-technology, or could be developed" into a practical tool. The same coverage floated a sharper worry: that reading deep meaning into a structural resemblance can be apophenia, the human habit of connecting dots that may not be connected. A shape that looks like CRISPR is a lead, not a discovery of a new gene-editing system.

Who actually did what

There is a second flattening in the phrase "AI discovered." Claude did the computational surfacing. The validation that the ART system is real, that the protein expresses and that the array produces distinct short RNAs, was done by human scientists in a lab, expressing the protein in standard strains and characterizing it biochemically and structurally. That is the correct division of labor, and it is more impressive stated honestly than dressed up as a solo machine discovery. The accurate headline is that an AI agent generated a strong lead from data at a scale humans cannot search, and humans confirmed the lead was worth something.

Feng Zhang, the MIT and Broad Institute scientist who helped turn CRISPR into a tool, landed on the right register. "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation," he said. Intriguing and worth investigating is exactly the claim the evidence supports. A new CRISPR is not.

Why it still matters

None of this is a debunking. Stripped of the marketing, what Anthropic showed is a clean example of AI doing real scientific work: an agent running for the better part of a day, over a dataset of genomic scale, surfacing a specific, checkable, previously-missed pattern that human experts then confirmed was really there. That is a template other labs will copy, and it is a more credible advertisement for AI in biology than any inflated "Claude discovered" headline, precisely because the honest version survives contact with independent experts.

It also sits awkwardly beside Anthropic's own safety branding, a company that warns loudly about AI-enabled biological risk now demonstrating how fast its model can mine biology at scale. That tension is worth watching, even if this particular enzyme turns out to do nothing at all.

Our take

Claude did something real and something oversold at the same time, and both are true. The real part: an autonomous agent surfaced a genuine, overlooked genomic pattern from 1.9 billion clusters in a day, and humans validated it. The oversold part: "CRISPR-like" describes a shape, not a function, and nobody, Anthropic included, can yet say what ART does or whether it is useful. Read this as a strong proof of concept for AI-assisted discovery, not as a new gene-editing breakthrough. The most impressive thing here is the method, and the method does not need the hype.