AI Makes a Genetics Discovery in Record Time

Two Documents, and Only One Got Read

"Claude discovered a CRISPR-like system in 21 hours." The line spread everywhere in about two days. It comes from Anthropic, which published a post on September 23 titled "Claude discovers a novel enzyme system with CRISPR-like repeats."

That evening, CEO Dario Amodei posted about it on X, minutes before addressing the UN Security Council. He did not oversell it: "its precise function, its biotech utility if any, and its level of importance are not yet clear." That quote comes from Gizmodo's writeup - X blocks automated reading, so we did not read it at the source.

But the announcement is two documents: a one-page post, and a forty-page preprint. Everyone read the first; the numbers live in the second, and they do not quite match.

A quick primer on CRISPR, since the story hinges on it. Bacteria keep snippets of DNA from the viruses that attack them, to recognize and cut them if they return; researchers turned that cutting mechanism into a gene-editing tool, work that won a 2020 Nobel Prize in Chemistry and reached an approved therapy by late 2023. What Anthropic describes is not that. It is a three-part assembly found in the DNA of viruses that infect bacteria - something that looks like CRISPR in shape, and whose purpose nobody yet knows.

Three Caveats, and Who Kept Them

Anthropic's own post carries three caveats. On function: "we don't yet know its function." On the enzyme at the system's core: it "had been identified by prior studies." And on publishing now: "to demonstrate Claude's capabilities."

Of eleven writeups we checked, eight kept those caveats. France's Numerama put the hedge right in its headline - "but isn't quite sure what it actually does" - and asked the French Genetics Society for comment. They declined: forty non-peer-reviewed pages don't get assessed in 24 hours. It is the most careful writeup of the bunch.

Four dropped the hedge, including one that promised a system "that could cut, copy, and paste." The phrase is not invented: it is in Anthropic's own post, describing a handful of other systems with similar traits, not ART. Moving it one sentence over was enough, the same way the 93 percent figure from the OpenAI lawsuit needed only that much to switch jobs entirely.

"In 21 Hours": 950 Workers, Not One

The preprint gives the numbers the blog post rounds off. The run lasted 21.5 hours on the clock - but it mobilized 949 separate agent sessions, totaling 77 cumulative agent-hours and 215.6 million tokens, with no human in the loop.

Those 21 hours are wall-clock time, not a measure of work: it is how long it took a crew of 950 to do what one worker would have needed three and a half days to finish. The post's comparison to "weeks to months" of human labor holds up, but it puts a wall of compute against a single person.

The trail is documented: 1.94 billion protein clusters searched, 198,290 kept after filtering, 3,564 related families flagged, 17 candidates examined closely. Of those 17, the preprint states only three were confirmed as associations never described before.

"CRISPR-like": the Architecture, Not the Function

Dimitri Perrin runs the computer science school at Queensland University of Technology and has led projects combining CRISPR and AI. His analysis in The Conversation boils down to one sentence: ART is CRISPR-like in its architecture, but nothing indicates it is CRISPR-like in what it does.

The preprint does not argue otherwise - it undercuts itself here. It notes that ART's repeats "contrast" with CRISPR's. More telling: no cas gene sits anywhere near an ART locus. Cas genes are the system's scissors; here, there are none.

ART's real neighbor is not CRISPR - it is the retron, a bacterial defense system that shares all three of its components. The figure comparing the two says so in its own caption: "retron (known) / ART (hypothesis)," with dotted arrows marking steps that are assumed, not shown.

The preprint's sharpest line is this one: the authors never demonstrated that the enzyme is active, or that these RNA strands are even what it acts on.

Anthropic leans on an endorsement from Feng Zhang, a pioneer of CRISPR gene editing and professor at MIT and the Broad Institute, who read the preprint. He calls the result "really intriguing" and says it is worth pursuing further. That is the extent of it - he never writes that ART edits genes.

"Discovered": Someone Had Already Walked Past It

"The enzyme itself was already known," Perrin writes. "What's new is recognizing it might be part of this larger system."

The preprint goes further. The original genome report on this stretch of DNA had already flagged the enzyme - and noted, right next to it, a fragment it never described. Someone had already been on the right page and missed the line next to it.

Lucas Harrington, from Jennifer Doudna's lab and a cofounder of Mammoth Biosciences, criticizes Anthropic for framing routine genome-mining work as a major discovery. Quoted by Numerama, he notes that this kind of search has been used to understand CRISPR since 2002.

That puts the timeline in perspective. The repeats were first spotted in 1987, in a paper about an entirely different question; the acronym CRISPR itself dates to 2002. Thirty-six years between the curiosity and the drug.

What the Preprint Says That the Press Release Doesn't

Perrin asks the question everyone skips: how many of the 3,500 candidates, and the 20 that made the cut, actually reflect real new biology? Finding one intriguing hit after combing through thousands, he writes, is not the same as showing a system discovers things reliably.

The answer sits in the document, on page 9. Anthropic reran the exact same campaign ten times, same setup, same prompt. Almost every run sampled ART loci. Two opened an investigation into the lineage. None read the DNA further upstream, and the repeat table was missed on every rerun.

Once out of eleven. The authors attribute it to the size of the search space and the non-deterministic nature of their setup, and they say so plainly. Nothing was hidden; it is right there, in the document almost nobody opens. The newsletter Run Data Run flagged it on September 25, noting that none of the coverage it had read mentioned it either. Neither had ours, until now.

The next section explains what actually decides the outcome, and it is counterintuitive. Better-equipped models look less: give them tools to manipulate sequences directly, and 39 percent of attempts never read enough DNA in a row to spot two repeats. Have the model read it itself instead, and the hit rate jumps from 29 to 76 percent.

It is the same gap as with Astra's 100 percent security score: the number that circulated is not the one that tells you anything.