Anthropic’s Claude Flags CRISPR-Like Repeats, Not a Tool

In a large genomic database, a repeating stretch of DNA beside a previously identified enzyme might be easy to overlook. Anthropic says Claude spotted one, prompting its scientists to investigate a biological system they now call array-associated reverse transcriptases, or ART.

Anthropic reports a laboratory result: the DNA array produces distinct short RNAs. That takes ART beyond an unexplained pattern in sequence data, though it does not establish what the system does, whether those RNAs guide the enzyme, or whether ART could become a gene-editing tool. The array resembles part of a CRISPR system, but resemblance is not a demonstrated capability.

Anthropic has released its account in a preprint linked from the company’s announcement. Its scientists conducted the reported experiments; independent replication is not established in the material available for this article.

Claude Found an Array Beside a Known Enzyme

Anthropic directed Claude to search genomic data for unusual reverse transcriptases. These enzymes copy RNA into DNA, and the genes around them can offer clues about the biological systems in which they operate.

According to Anthropic’s account of the search, roughly 950 Claude agents worked over 21 hours, using about 210 million tokens. They gathered more than 200,000 reverse transcriptases, identified approximately 3,500 candidate systems, and narrowed those to 20 for closer analysis. During that work, an agent noticed evenly spaced DNA repeats near the gene for an unusual reverse transcriptase.

The agent examined the repeats’ number and spacing, compared the arrangement with known systems, and searched for previous descriptions before submitting a report for human review, Anthropic says. Its scientists then investigated the candidate.

Anthropic says earlier research had already identified the underlying reverse transcriptase in a jumbo bacteriophage. Claude appears to have flagged a previously uncharacterized arrangement around it: the enzyme gene, a neighboring gene for a partner protein of unknown function, and a long array of repeated DNA. Bacteriophages are viruses that infect bacteria; Anthropic says ART systems occur mainly in them.

Sequence databases contain far more genes than researchers have characterized experimentally. A familiar-looking protein can belong to an unfamiliar system if its genomic neighbors reveal a pattern no one has connected before. The scale of Claude’s search helps explain how it found a candidate, but the agent count and token total say nothing by themselves about whether the proposed biological interpretation is right.

The Lab Found Short RNAs, Not a Known Function

A repeat array in DNA is an observation. Researchers need experimental evidence to find out whether it participates in an active system. Anthropic says its scientists tested the ART candidate and found that the array is expressed as a set of distinct short RNAs.

The repeated sequence is therefore producing RNA molecules, not merely sitting beside the enzyme gene. Those molecules might be involved in the system’s activity, but expression does not tell researchers what they do. Anthropic has not established ART’s primary function or shown that the reverse transcriptase uses the short RNAs for a particular operation.

Claude conducted the database search and candidate analysis after receiving a high-level research prompt, according to Anthropic. People reviewed the finding and performed all physical laboratory work. The company describes an ongoing workflow in which its scientists assess AI-generated candidates and use laboratory results to decide what warrants further study. This was an AI-assisted discovery process, with human scientists running the experiment and interpreting its results.

The preprint described in Anthropic’s announcement provides the team’s early research account. A preprint makes findings available before journal peer review; it does not amount to independent reproduction of the reported lab result. For now, Anthropic reports a genomic arrangement that its scientists followed up with evidence of RNA expression.

CRISPR-Like Describes the Pattern, Not the Capability

The CRISPR comparison concerns the layout of the DNA repeats and the production of short RNAs. In established CRISPR-Cas systems, arrays provide sequences that help guide associated proteins toward targets. That relationship between RNA and protein is central to why CRISPR can be adapted for targeted editing.

Anthropic has not demonstrated that ART’s short RNAs guide the reverse transcriptase to a chosen sequence, that the system edits a genome, or even what biological task it performs. A reverse transcriptase copies RNA into DNA; finding one beside an RNA-producing array does not establish the targeting and editing behavior of a CRISPR tool.

Anthropic says the combination of features reminds its researchers of a small set of other programmable systems. The comparison gives them a reason to investigate ART without showing that it shares those systems’ functions. The company’s own announcement says work to determine ART’s primary function is ongoing. As The Verge’s coverage notes, whether the finding will have a practical application remains unclear.

“Claude invented a new CRISPR tool” misstates the result. Claude did not design the enzyme, and Anthropic has not shown a programmable editor. It reportedly recognized a potentially meaningful association in existing genomic data, which human scientists then began to test.

A Discovery Claim With Two Separate Tests

Two questions remain: Is ART a distinct, previously uncharacterized biological system, and how much credit does Claude’s search process deserve for identifying it? Anthropic’s published account supports examining both, but does not settle every part of either one.

Other researchers can scrutinize the reported sequence arrangement and test the RNA result independently. Establishing the mechanism would require finding out how, or whether, the short RNAs interact with the enzyme and its partner protein. Without that work, researchers cannot say what ART does in a phage, much less assess its potential as a biotechnology tool.

The most concrete evidence for the AI claim is the described trail from a broad reverse-transcriptase search to an agent’s report and a human-run experiment. Anthropic chose the research question, built the workflow, reviewed candidates, and carried out the lab work. Calling the result entirely autonomous would obscure those decisions. Dismissing it as ordinary pattern matching would miss the reported ability to search, compare and prioritize an anomaly worth testing.

A Hacker News discussion of the announcement reflected that tension. Some commenters argued that “novel enzyme system” sounded stronger than the evidence for a previously undescribed arrangement around a known enzyme; others considered that arrangement worth investigating even without a known function. Those comments are reactions, not scientific verification. The distinction they raised is useful regardless of where a reader stands on AI’s contribution.

Final Thoughts

ART offers a concrete example of what an AI research assistant can contribute: a candidate pulled from a large sequence search, followed by a human experiment that found something measurable. The short RNAs make the arrangement worth studying, while leaving its function open.

Functional evidence, ideally tested beyond Anthropic’s lab, would be the next decisive result. Until then, the achievement to evaluate is the identification and initial investigation of an overlooked biological arrangement, not the arrival of another CRISPR editor.