Claude found a CRISPR-like enzyme system — but nobody knows what it does
Anthropic announced that its AI model Claude autonomously identified a previously uncharacterized enzyme system in bacterial DNA, with a structure reminiscent of CRISPR. The findings have been published as a technical report, not peer-reviewed, and experts are already publicly divided about what they actually mean.
The news
Anthropic published the findings on its website on September 23, 2026, and promoted them more broadly on Wednesday, September 24 — alongside the launch of a new life sciences research group and a wet laboratory in the Bay Area that opened in the spring. According to The New York Times, this is the first result from the new laboratory.
The company's founder and CEO, Dario Amodei, wrote on X that the company suspects the "molecular machine" may "represent a new gene editing mechanism" (Al Jazeera). Amodei has previously claimed that AI could cure most diseases within five to ten years.
It is important to keep the facts straight: Anthropic itself has stated that the function of the system has not been determined. What is new is not the enzyme itself, but the pattern surrounding it.
What ART is
Anthropic calls the three-part setup "array-associated reverse transcriptases," abbreviated ART. The system consists of three parts: a reverse transcriptase enzyme, a neighboring gene whose function is unknown, and a long array of evenly spaced DNA repeats (Yahoo Science).
Reverse transcriptases are not new — enzymes of this type were already catalogued. What Claude reportedly found was the surrounding pattern that researchers had overlooked: the combination of the enzyme, the partner protein, and the repeat array, which together form a coherent system.
According to the Yahoo report, the discovering agent logged its own surprise when it noticed the repeat array in raw DNA.
The campaign's mechanics — as Anthropic tells it
The figures below come from Anthropic itself, relayed through secondary coverage, and have not been independently verified:
- Approximately 950 Claude agents worked in parallel on the same search, launched by Anthropic's researchers with a single instruction: find novel reverse transcriptases in a massive DNA database.
- The search ran for 21 hours.
- The agents used 210 million tokens — small units of text that AI models process. (TechTimes reports 215 million in its headline; the majority of sources give 210 million, and the discrepancy has not been clarified.)
- The agents collected more than 200,000 reverse transcriptases, picked out 3,500 novel candidate systems, and narrowed the field to the 20 most convincing candidates.
One of those 20 led, then, to ART.
What the wet-lab work shows so far
According to the preprint, as reported by TechTimes, Anthropic's initial laboratory experiments showed that the ART array is transcribed into a set of distinct short RNAs. In publicly available data from a Staphylococcus phage, these RNAs accounted for up to 8 percent of total phage RNA 15 minutes after infection.
This means the array is biologically active in bacteriophages — but it does not yet say what the system does, or whether it can be programmed the way CRISPR can.
The experts are divided
Reaction from the biology community spans the full range:
Feng Zhang, CRISPR pioneer at the Broad Institute, has reviewed the preprint and calls the finding "an exciting example of how AI agents can contribute to biological discovery," saying that the identification of RNA repeat arrays linked to reverse transcriptases is "genuinely intriguing and merits further investigation" (TechTimes). But he claims neither that ART edits genes nor that it will become the next CRISPR.
Stanley Qi, associate professor of bioengineering at Stanford, describes the report as "incredibly exciting," highlighting the model's ability to recognize unusual biological patterns (Al Jazeera).
On the defensive side, Philip Kranzusch, a microbiologist at Harvard Medical School, told The New York Times: "It's not yet a breakthrough. It's a wrinkle on what we already know in the field."
Kevin Blake, a microbiologist at Washington University School of Medicine, was even more pointed: "There's nothing to suggest this is a rival to CRISPR-as-a-technology, or that it can be developed into any kind of therapeutic or practical application" (Al Jazeera).
A Stanford group has previously identified a different system with some similarities to what Claude found, suggesting such structures may be sitting waiting in the databases (RedState).
It is worth noting the distinction between two different claims: that Claude found a real, uncharacterized biological pattern (which several experts appear to accept), and that this constitutes a new gene editing tool (which no one can currently demonstrate).
The market reacted quickly
Shares in gene therapy companies fell on Wednesday after the announcement. Beam Therapeutics, CRISPR Therapeutics, Intellia Therapeutics, and Prime Medicine were among the companies investors weighed, according to Yahoo Finance, in light of the potential implications for gene editing research.
The reaction says something about how the market interpreted the message — but it says nothing about whether ART will ever become clinically relevant. The function is unknown, and no one has demonstrated therapeutic applicability.
Biosafety and strategic context
Anthropic was upfront with safety information about the new laboratory: it is located in the Bay Area, resembles a typical molecular biology lab, operates only at the lower biosafety levels (BSL-1 and BSL-2), does not handle pathogens that can infect humans, and all lab work is performed by human researchers — not robots controlled by AI (Anthropic, cited via RedState).
Eric Kauderer-Abrams, head of life sciences at Anthropic, explained the company's bet to The New York Times: "We've chosen biology and medicine as the primary pathway we think the benefits from AI will come through."
The announcement comes from a company that has been central to the global debate about how to maximize the benefits of AI while mitigating catastrophic risks — a tension several commentators have pointed to between Anthropic's warnings and now this positive demonstration.
What remains
Three things must fall into place before anyone can call ART a new CRISPR:
- The function must be determined. Anthropic has not yet managed that itself. Nothing so far suggests the system can be programmed to edit DNA.
- Peer review. The results exist only as a technical report on the company's own website, not submitted to any journal (The New York Times).
- Independent verification. The campaign figures — 950 agents, 21 hours, 210 million tokens — are Anthropic's own claims. Kranzusch's point that this is "a wrinkle on what we know" reflects that reverse transcriptases were already known.
What has actually been demonstrated, for now, is something different and less dramatic: that a swarm of AI agents can search through enormous DNA databases, find a structural pattern the professionals had overlooked, and deliver it to the laboratory for further testing. Whether that pattern turns out to be a tool or merely a biological curiosity, nobody knows today.