Anthropic’s Claude Found a CRISPR-Like DNA System. Scientists Say the Real Test Is Just Beginning

Anthropic says its Claude AI models helped uncover a previously unrecognized DNA-related system in jumbo phages, a finding the company compared to CRISPR because of its apparent genome-cutting potential. The discovery, announced on September 23, could become important for gene editing research, but scientists say it is still only an early lead that has to be proven in the lab.

The claim matters because it is one of the clearest examples yet of an AI system surfacing a biological pattern that researchers think may point to a useful tool. It is also a reminder that pattern-finding is not the same as scientific validation: the discovery is promising, but it remains unreviewed, partly tested and far from any practical application.

What Anthropic says Claude found

Anthropic says Claude identified an unusual enzyme system that appears related to reverse transcriptases and includes a long stretch of repeated DNA sequences that resemble the structure of CRISPR arrays. The company has named the system ART, short for array-associated reverse transcriptases.

According to Anthropic, the model’s output pointed researchers toward a genomic pattern found in jumbo phages, large viruses that infect bacteria. The company says the system was uncovered after Claude agents searched massive genomic databases for unusual examples of reverse transcriptases, enzymes that copy RNA into DNA.

Anthropic described the finding as a discovery with “properties reminiscent of CRISPR,” the gene-editing technology that transformed biology and medicine after years of basic research. But the company also acknowledged that it does not yet know what the system does or whether it can be turned into a gene-editing tool.

Anthropic said the system shares features with only a small number of known mechanisms, all of which can cut, copy or move DNA, but the company stressed that its function is not yet understood.

How did Claude reach the discovery?

Claude did not arrive at the finding in one step. Anthropic says about 950 Claude agents ran at the same time and completed the search in 21.5 hours, scanning huge datasets and narrowing down a field that initially included more than 200,000 possible reverse transcriptases.

The process was essentially a large-scale filtering exercise. Claude first identified a broad universe of candidate enzymes, then reduced that set to several thousand that appeared novel, and finally singled out a smaller family with a repeat-rich region that looked unusual enough to merit attention from researchers.

One of the AI agents even flagged the pattern in informal terms, describing a tandem repeat array that looked “CRISPR-like.” The technical report also noted the possibility that the system could resemble a retron, another bacterial immune mechanism that can sometimes be adapted for editing, though it is not equivalent to CRISPR.

Why the database search mattered

The project highlights where large language models may be most useful in biology: not as autonomous scientists, but as high-speed pattern detectors across enormous datasets. Mining genomic databases by hand can take months, especially when the target is a rare or unfamiliar enzyme family.

By that standard, Anthropic’s claim is striking. The company says the search, which would normally require substantial manual effort, took less than a day once Claude agents were deployed. That speed is what impressed some scientists, even those who are skeptical of the broader claim.

Why are scientists cautious about the CRISPR comparison?

Scientists caution that a CRISPR-like pattern is not the same thing as a CRISPR system. The feature that made CRISPR revolutionary was not merely that it looked unusual, but that researchers eventually showed it could be programmed into a precise and versatile genome-editing tool.

That leap has not happened here. Anthropic included only one physical experiment in a technical report, and that report has not been peer-reviewed. Researchers will need to determine whether ART really has editing activity, how it works, and whether it is useful for any real-world application.

The comparison to CRISPR may therefore be more rhetorical than scientific at this stage. The newly identified system may turn out to be a retron-like mechanism, an interesting biological finding in its own right but not necessarily a new editing platform.

Le Cong, a Stanford professor studying AI in genome engineering, argued that the key issue is not the announcement itself but the experimental backlog that still has to be cleared.

Cong’s point was blunt: the model may have found something intriguing, but the hard work of validation still sits in the lab. In his view, the public relations side of the story moved faster than the science.

What do outside experts think?

Reaction from researchers has been mixed, with enthusiasm about the method but skepticism about the claim that AI independently made a meaningful biological discovery.

Fyodor Urnov, a gene-editing specialist at the University of California, Berkeley and a director at the Innovative Genomics Institute, praised Anthropic for making the finding public. He noted that the institute collaborates with Anthropic, although it was not involved in this specific work.

Other experts were less impressed with the framing. Seth Shipman, an associate investigator at the Gladstone Institutes, said the more interesting part is the workflow that surfaced the pattern, not the biological system itself. He also urged caution in describing the result as something Claude discovered on its own, since humans designed the prompts, set the search goals and interpreted the output.

Shipman said the underlying enzyme family was not entirely new, but the AI-driven search made it much faster to spot relevant surrounding features that might hint at a CRISPR-like system.

Jason Gill, a microbiologist at Texas A&M University, added another layer of nuance. He and colleagues had already identified the same reverse transcriptase in a 2021 jumbo-phage paper. What may be new, he said, is that Claude spotted repeat sequences around it that suggested a larger system architecture.

Gill emphasized that models can be excellent at pattern recognition, sometimes better than humans looking at a screen, but they still need a scientist with a hypothesis to know what the pattern means.

What is ART, and how is it different from CRISPR?

ART is Anthropic’s name for a reverse-transcriptase-associated repeat system discovered in jumbo phages. Reverse transcriptases are enzymes that convert RNA into DNA, the reverse of the usual cellular flow of genetic information.

CRISPR, by contrast, is a bacterial immune system repurposed into a programmable editing tool. It relies on guide molecules and DNA-cutting enzymes to target specific sequences. Retrons, another type of bacterial system mentioned by researchers, can sometimes be used in gene editing but are not the same as CRISPR.

At this stage, ART is best understood as a candidate biological system with some intriguing structural similarities to known DNA-manipulating systems. Whether those similarities translate into useful function remains unknown.

How much of this is really an AI discovery?

That depends on how one defines discovery. Anthropic’s system appears to have done the heavy lifting of sorting through enormous datasets and surfacing a strong candidate. But the experiment was designed by researchers, the search problem was chosen by humans, and the interpretation still depends on human expertise.

This is why several scientists are pushing back on the idea of fully autonomous scientific invention. In their view, Claude acted more like a powerful research assistant than a self-directed scientist.

The distinction matters because the phrase “AI discovered X” can imply a level of independence and reliability that does not yet exist. The current reality is messier: AI can accelerate one part of the scientific pipeline, but it cannot replace validation, context or experimental design.

Could Claude have been trained on the answer already?

That question is difficult to answer from the outside. One concern raised by researchers is that the model’s training data are not fully transparent to the scientific community, which makes reproducibility harder than in conventional academic research.

In a related dispute, one scientist who had been studying similar enzymes became suspicious that Claude may have seen his unpublished work after he discussed it publicly with the model. Anthropic has said Claude was not trained on user transcripts and that its molecular biology team does not have access to them either.

Even with that reassurance, the episode highlights a broader tension around AI-assisted science: researchers want speed and scale, but they also need clarity about what data the model used and how it arrived at its conclusions.

Why this discovery matters for drug discovery and biology

The finding matters because biology increasingly depends on sifting through huge molecular archives that are too large for manual review alone. If models like Claude can consistently surface credible candidates, they could shorten the time from data collection to hypothesis generation.

That would be especially valuable in fields such as gene editing, enzyme engineering, antibiotic discovery and synthetic biology, where the search space is vast and the payoff from a single useful mechanism can be enormous.

But the possible upside should not obscure the limits. A promising sequence does not become a medicine, a therapy or even a validated tool without a chain of follow-up work that remains stubbornly human.

What has to happen next?

First, researchers need to confirm whether ART is actually functional. That means more experiments, better characterization and peer review. Then they must determine whether the system can be controlled or adapted for editing.

After that comes the harder question: even if ART is active, is it useful? Many biological systems are interesting on paper but too inefficient, too narrow or too difficult to engineer for practical use.

• Replication in independent labs
• Peer-reviewed publication of the data
• Functional assays to test editing activity
• Engineering studies to assess usability
• Safety and biosafety review for any future applications

How does this compare with the history of CRISPR?

The comparison is tempting because CRISPR itself began as a basic-science curiosity before becoming one of the most important tools in modern biology. But that transformation took decades, not weeks or months.

Researchers first noted the unusual bacterial sequences in the late 1980s. It was not until 2012 that Jennifer Doudna and Emmanuelle Charpentier showed CRISPR could be programmed as a gene-editing tool. Even then, clinical and commercial translation has been slow and tightly regulated.

There is now only one approved drug on the market that uses CRISPR technology, despite dozens of clinical trials across blood disorders, cancer, autoimmune disease, cardiovascular conditions and other areas. That history is a cautionary tale: basic discoveries can be revolutionary, but the path to medicine is long and unpredictable.

What Anthropic’s CEO is signaling

Anthropic chief executive Dario Amodei used the announcement to point toward a future in which AI could do more than search databases. He suggested that one day Claude might directly operate lab equipment and run experiments on its own.

That vision remains distant, and it raises immediate concerns about safety, supervision and biosafety. Amodei later noted that Anthropic’s labs operate at the lowest biosafety levels, which means they are not set up to handle especially dangerous biological materials.

The broader implication is clear: Anthropic wants to present its models not just as text generators, but as research tools with scientific utility. Whether the company can convince the biology community will depend on results that can be checked, repeated and independently confirmed.

What this means for AI in science

The most defensible takeaway from Anthropic’s claim is not that AI has replaced scientists, but that it may already be capable of accelerating parts of scientific discovery that involve huge search spaces and subtle patterns. That could change the pace of hypothesis generation in many fields.

At the same time, the case shows why AI-driven science still needs skepticism. Models can point to promising structures, but they do not automatically know what is biologically meaningful. They can also obscure how a result was produced if the data and methods are not transparent enough for independent review.

In that sense, ART is less a finished breakthrough than a test case. It may become an example of AI-assisted biological discovery done well. Or it may turn out to be a useful hint that never grows into anything practical.

Either way, the announcement has pushed a larger question into the open: can a model trained on vast prior knowledge help find what humans have not yet recognized, and do so in a way science can trust? For now, the honest answer is maybe—but only after the lab work catches up.

Timeline of the ART claim

1. 2021: Researchers including Jason Gill identify the reverse transcriptase in jumbo phages.
2. Earlier in 2026: Anthropic forms a research group and sets up a wet lab for drug discovery.
3. September 23, 2026: Anthropic publicly announces Claude’s search result and names the system ART.
4. After the announcement: Scientists emphasize that the finding needs more experiments, peer review and independent confirmation.

The bottom line

Anthropic has shown that AI can rapidly sift enormous biological databases and surface an intriguing candidate that might one day matter for gene editing. But the company has not yet shown that Claude discovered a new CRISPR system, nor that ART has any practical use.

For now, the finding is best understood as a promising lead with real scientific curiosity attached to it. The real verdict will come from the lab, not the model.