Did 950 AI Agents Really Discover a New CRISPR-Like System?

It had been reading raw DNA beside an unusual enzyme gene when a pattern emerged: the same short sequence repeating again and again, separated by longer stretches of genetic text. To the model, the arrangement looked uncannily familiar. “That’s a CRISPR-like … repeat array?!” it wrote in its work log.

The exclamation would become the cinematic center of a claim from Anthropic, the company behind Claude: approximately 950 AI-agent sessions had searched an immense genomic database and surfaced what the company describes as a previously uncharacterized biological system. The system, found mainly in giant viruses that infect bacteria, was named ART—array-associated reverse transcriptases.

Anthropic says its AI did in roughly a day what could take expert researchers weeks or months. Then the scientists began arguing.

Some researchers called the pattern intriguing and the AI workflow impressive. Others objected to the CRISPR comparison, noted that the underlying enzyme had appeared in earlier research and warned that the system’s function remains unknown. A parallel controversy emerged over what “autonomous discovery” means when humans selected the problem, built the tools, reviewed the output and performed every physical experiment.

The result is more interesting than either the hype or the backlash. Claude appears to have detected a meaningful biological anomaly at extraordinary scale. But nobody has demonstrated that ART edits genes, defends bacteria, behaves like CRISPR or has any practical use.

It is the story of science encountering a new kind of participant—and discovering that we do not yet have rules for deciding how much credit it deserves.

1. AI Found a Strange DNA Pattern

Anthropic’s technical preprint on autonomous genome mining asked whether language-model agents could behave more like tireless scientific curators. Instead of only applying fixed filters, agents could write code, use bioinformatics tools, inspect raw sequences, read literature, form hypotheses and open follow-up investigations when something looked unusual.

According to the preprint, the campaign searched 1.9 billion protein clusters, recovered roughly 200,000 RT clusters, classified them into nine groups and examined thousands of recurring neighboring protein families. It ultimately produced 19 candidate reports.

At some point during a 21.5-hour digital hunt, one AI agent stopped. It had been reading raw DNA beside an unusual enzyme gene when a pattern emerged: the same short sequence repeating again and again, separated by longer stretches of genetic text. To the model, the arrangement looked uncannily familiar. “That’s a CRISPR-like … repeat array?!” it wrote in its work log.

The AI had not been explicitly directed to search for this architecture. It was investigating protein partners when it inspected the surrounding DNA and flagged the repeated pattern.

2. AI Identified An Unusual DNA Repeat Pattern

The surprising candidate appeared beside a reverse transcriptase associated with jumbo phages—viruses with unusually large genomes that infect bacteria.

Upstream of the RT gene was a long stretch of non-protein-coding DNA arranged into tandem repeats. In the representative sequence highlighted by the researchers, a short core motif recurred across multiple units separated by variable sequences.

Anthropic’s researchers call the system array-associated reverse transcriptase, or ART. It contains three central elements: the reverse transcriptase, a neighboring partner gene of unknown function and a long array of evenly spaced DNA repeat units.

3. The CRISPR-Like Pattern

The resemblance to CRISPR is architectural. Natural CRISPR systems contain arrays of repeated sequences separated by variable “spacers,” many derived from viruses encountered by bacteria or archaea. The array can be transcribed into guide RNAs that help CRISPR-associated proteins recognize matching invaders.

ART also appears to produce multiple short RNAs. In laboratory analyses of Staphylococcus phage infection, the researchers report that ART arrays were highly expressed and appeared as discrete units. That supports the idea that the repeated DNA is functional rather than meaningless genomic clutter.

The preprint does not show ART cutting a target sequence. It does not demonstrate memory of past infections, programmable recognition, bacterial immunity or gene editing. The researchers do not yet know what the system’s primary biological function is. “CRISPR-like” therefore describes the arrangement of repeats and the possibility of an RNA repertoire—not a proven mechanism.

4. What Does Art Actually Do

ART could represent another variation on this theme. An array capable of producing many distinct RNAs beside an RT raises an obvious possibility: perhaps the enzyme copies those RNAs into a diverse library of DNA products. The partner protein might direct, regulate or execute whatever comes next.

The system might defend a virus against competing viruses or bacterial countermeasures. It might regulate phage genes. It might generate molecular diversity. It might be a form of retron rather than an entirely separate category. Until researchers identify the products, targets and biological consequence, ART is a provocative arrangement with an unknown job.

5. AI Found It, Scientists Test It

The result is more interesting than either the hype or the backlash. Claude appears to have detected a meaningful biological anomaly at extraordinary scale. But nobody has demonstrated that ART edits genes, defends bacteria, behaves like CRISPR or has any practical use.

ART is at the earliest stage: an architecture, expression evidence and a mechanistic question. To justify comparisons with programmable gene editing, researchers would need to determine what RNA and DNA products ART creates, identify the partner protein’s role, establish its natural function and show that the system can be directed toward chosen targets. They would also need to test efficiency, specificity, delivery, toxicity and compatibility with different cell types.

6. New System Or New Interpretation

Anthropic acknowledges that the underlying jumbo-phage reverse transcriptase had been identified in previous studies. The company argues that Claude’s contribution was recognizing a broader system: the RT together with the repeat array and accessory protein.

Wired reported that the same RT appeared in a 2021 paper on jumbo phages. Critics therefore question language suggesting Claude found something wholly unknown. Seth Shipman, a retron researcher at the Gladstone Institutes, told Wired that he did not think ART was a new CRISPR system, although he considered the method of finding it interesting.

This distinction is familiar in science. Finding an object and recognizing what it belongs to are different discoveries. Astronomers may have imaged a point of light years before someone identifies it as a planet. A gene may sit in a database before researchers notice it participates in a previously undescribed pathway.

The fair claim is narrow: the RT itself was not new, but the Anthropic team says the repeated non-coding array, partner gene and RNA expression together define a system that had not previously been characterized as such.

The paper is also a preprint that has not completed peer review. Its methods, novelty claims and interpretation have not yet passed the scrutiny of journal referees. Anthropic released extensive agent transcripts and technical details, which helps outsiders evaluate the workflow, but transparency does not substitute for replication.

7. Not The Next CRISPR — Yet

This distinction is familiar in science. Finding an object and recognizing what it belongs to are different discoveries. Astronomers may have imaged a point of light years before someone identifies it as a planet. A gene may sit in a database before researchers notice it participates in a previously undescribed pathway.

The fair claim is narrow: the RT itself was not new, but the Anthropic team says the repeated non-coding array, partner gene and RNA expression together define a system that had not previously been characterized as such. Whether that system is biologically distinct will depend on experimental function and independent review.

The first draft of a new scientific contract

If agents can originate observations, what evidence must accompany an AI-discovery claim? How much human direction disqualifies “autonomous”? Who receives credit when a model notices a pattern using tools, databases and concepts built by generations of scientists? How should companies establish novelty when their systems may have absorbed vast amounts of public knowledge?

The answers cannot come from AI companies alone. Journals, funders, universities, database maintainers and working scientists will need standards.

A credible AI-discovery paper should disclose the model and version, initial instructions, accessible data, tools, intervention points, agent logs, human decisions, failed candidates and experimental validation. Independent groups should be able to reproduce the computational trail and test the biology.

ART may become a biotechnology platform, an interesting phage mechanism or a footnote. The more durable discovery may be procedural: AI can now do enough of science that science must decide how to audit AI.

The bigger question is whether AI can move from solving problems to making genuine discoveries—a question explored in our AI Solved a $1 Million Math Problem-But Did It Actually Discover Anything New?