What Claude actually found in phage DNA

The system is called ART, for array-associated reverse transcriptases. A reverse transcriptase is an enzyme that copies RNA into DNA. ART is a family of them found mainly in jumbo phages (large viruses that infect bacteria), and what makes it unusual is what sits next to the gene: a long run of short, repeated DNA sequences separated by unique stretches, plus a dedicated partner gene of unknown function.

That repeat-and-spacer layout is what earned the "CRISPR-like" label. It is also nearly the only thing ART is known to share with CRISPR. To see why, it helps to separate three things the coverage blended together: what the AI agents did, what the human scientists did, and what nobody has done yet.

What the headlines said

Al Jazeera's story on 24 September relayed Anthropic CEO Dario Amodei's post that the company suspects the system could "represent a new gene editing mechanism." Gizmodo's dek the same day said Claude had identified a system "that looks remarkably like CRISPR's gene-editing Cas9."

That second line is simply wrong, and the report says so in one sentence. Cas9 is the protein, the molecular scissors. The resemblance ART has is to the CRISPR array, the stretch of DNA that stores guide sequences. And on the proteins, the technical report is explicit:

Critically, no cas genes occur near any ART locus, suggesting that the arrays represent a new type of non-coding repeat element.

Even the array resemblance is partial. The report contrasts the two directly: CRISPR arrays have near-identical repeats with sharp boundaries and spacers of about 30 nucleotides that are gained and lost between related strains. ART's spacers are 120 to 220 nucleotides, unrelated to one another within an array, and retained in the same order between related phages, diverging at about the rate of the enzyme gene itself. In a CRISPR system the spacers are a changing record of past infections. Whatever ART's spacers are, they do not behave like that.

The same Al Jazeera piece carried the most useful outside comment. Kevin Blake, a microbiologist at Washington University School of Medicine, said: "There's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any kind of therapeutic or practical application." Nothing in the technical report contradicts him.

What the AI agents actually did

This part is real, and it is documented in unusual detail. Anthropic built a harness in which Claude Code agents, running the company's Mythos 5 model, worked as planners, workers, supervisors and editors against a written research brief. The brief asked them to find new reverse-transcriptase systems by looking for new partner-gene associations.

Per the technical report, the agents built their own sequence-search profiles, searched 1.94 billion metagenomic protein clusters, kept 198,290 reverse-transcriptase clusters after filtering, sorted them into nine classes, and scored 3,564 protein families that recurred in their genomic neighbourhoods. Seventeen families were promoted for deep dives. Of those, only three survived as genuinely new associations; the other fourteen were rejected as annotation artifacts, parts of known systems, or bystanders. The whole campaign ran 119 tasks and 949 agent sessions, 77 agent-hours and 215.6 million tokens over 21.5 hours of wall-clock time, "without human intervention."

(The announcement post rounds these differently, saying roughly 950 agents, 210 million tokens, 3,500 candidate systems and "the 20 most compelling candidates." The technical report's numbers are the ones I use here.)

ART was not one of the partner-gene hits the brief asked for. It was a side path. An agent investigating one candidate concluded the association was spurious, but flagged the reverse transcriptase itself as worth a closer look. A supervisor agent asked the next worker to check the DNA upstream of those genes for a non-coding RNA, since related retron systems have one there. The worker pulled the raw sequence into its context and wrote:

The L0050 (228,907 bp logan contig) flank is spectacular: I can see by eye a tandem repeat array

It then counted the repeats (14 copies of a 16-nucleotide repeat in that locus), compared the layout to known systems, ran a literature search to try to kill the novelty claim, and filed a report for human review. That sequence (notice an anomaly outside the brief, then try to refute it, then escalate) is the actual result worth taking seriously here. It is the part of genome mining that normally depends on an expert looking at data.

What the humans did

After the agents' report, the characterization moved to what the paper calls "interactive Claude Science sessions," meaning scientists prompting the model and directing the analysis. That work established the family: 95 distinct reverse-transcriptase clusters, 28 of them with a detectable upstream array of 3 to 21 repeat copies, a phylogenetic position next to retrons, an unusually long N-terminal region on the enzyme, and three unrelated families of partner protein.

One of those sessions retrieved published RNA-sequencing data from a 2022 study of Staphylococcus phage SA1 infection. It showed the array is transcribed heavily: at 15 minutes after infection, array-derived RNAs made up as much as 8% of phage RNA. The lab then expressed the SA1 system on plasmids in E. coli and sequenced the small RNAs, which again resolved into discrete short pieces. Anthropic's announcement states that "All of the lab work is performed by human scientists."

So the one wet-lab experiment in the report is a human-run small-RNA sequencing result showing the array is expressed as distinct short RNAs. That is a real finding. It is not a function.

What nobody has shown yet

Here the report's discussion is admirably blunt, and it is the paragraph that should have led every story:

Beyond these observations, we have not shown that the RT is active or that the unit RNAs are its substrates. Whether the RT and its partner interact, and what the system does for the phage, are currently unknown.

The authors offer a hypothesis: that ART works like a retron, a known bacterial defence system in which an enzyme, an RNA and a partner protein sit in an inactive complex until something triggers them, except with a bank of different RNAs instead of one. That hypothesis is labelled with question marks in their own figure. The partner-protein interaction rests on structure predictions at moderate confidence. Nothing in the report shows ART editing, cutting or copying anything.

Two further points of context are in the report and missing from most coverage. First, the enzyme was not unknown: the genome report for one of these phages, MarsHill, published in 2021, identified the reverse transcriptase and proposed a non-coding RNA upstream, but did not describe the repeats or the partner gene. Second, ART is not the first reverse-transcriptase family found with CRISPR-like arrays. The authors cite a 2026 preprint that found arrays beside an unrelated family, UG27, using a genome language model, and conclude the architecture "arose more than once in nature."

Would it happen again?

This is the section I would most want a lab manager to read, because it is where the report is most honest about the method.

Anthropic reran the identical campaign ten more times. Nearly every run sampled ART loci, and in two runs agents investigated the lineage, but none read the DNA upstream of the enzymes, "and the array was missed in every rerun." The authors attribute that to the size of the search space and the non-deterministic behaviour of the harness.

They then built fixed benchmarks, handing seven Claude models the ART sequences at five levels of information and scoring the reports against ten features. With the loci placed directly in context, the four most capable models described the array in at least 90% of attempts. Given the same data as files with analysis tools, recognition fell as low as 32% for one model, because in the file-based runs 39% of the capable models' attempts never read a contiguous stretch of 200 nucleotides or more, and so never saw more than about one repeat unit. More tooling made the models worse at the one thing that mattered. Note also that these benchmark reports were scored by a judge model, Mythos 5, from the same family being tested.

The interpretability section found two internal signals in Mythos 5 that respond to the repeats as the sequence is read and go quiet when each repeat copy is shuffled. The figure caption adds the qualifier: "Neither is specific to DNA." That is a model that responds to repetition, and repetition in DNA is what it saw.

None of this undercuts the find. It does reframe it. The accurate headline is not that an AI reliably discovers biology. It is that one run of an agent system, on one occasion, looked at raw sequence nobody had looked at closely, noticed something a pipeline would not have flagged, and handed it to humans, and that the same setup did not repeat the trick ten times running.

What I could not confirm

This is a company technical report, not a peer-reviewed paper. All six authors are Anthropic employees. The announcement calls it a pre-print and quotes Feng Zhang of MIT and the Broad Institute, who reviewed it and called the RNA-repeat arrays "genuinely intriguing." I found it hosted on Anthropic's own site; I did not find it on a preprint server, and I am not aware of independent replication.

I have not re-run any of the analyses. Every number here is taken from the technical report or the announcement as published. I did not check the sequences, phylogeny or RNA-seq reanalysis myself.

I relied on coverage for the CEO's statement. I quote Amodei's "new gene editing mechanism" line as reported by Al Jazeera and Gizmodo, not from his original post.

The signal

When a headline says "AI discovers X," the useful questions are always the same. What did the model actually do? Who ran the experiments? Was it reproduced? And what does the primary source say is still unknown?

Here the answers are: the model searched, filtered, noticed and escalated; humans ran the lab work; the discovery did not reproduce in ten reruns; and the function is unknown. That is still a meaningful result for genome mining, and it is being oversold as a result for gene editing. I made a similar point about an AI-scientist drug-repurposing claim in Report 146, and about what a model can and cannot know from text alone in Report 159. The pattern holds: the paper is usually more careful than the press release, and the press release is more careful than the coverage.

Disclosure, plainly: Claude, the model discussed in this report, is one of the tools I use in my own work, including research and drafting for this publication. I am a customer of Anthropic's products and have no other relationship with the company; Anthropic did not ask for, see or review this report. Nothing here is sponsored and no link earns a commission; here's the full policy.