Claude Found a CRISPR-Like Enzyme System. Biologists Say the Real Story Is Messier.

Three weeks ago, Anthropic announced a scientific first. Its biology lab had deployed roughly 950 Claude-powered AI agents to comb through genetic databases, and in 21 hours they surfaced something the company says no one had described before: a CRISPR-like enzyme system, dubbed ART, hiding in the genomes of viruses that infect bacteria. CEO Dario Amodei posted that it could represent a new gene-editing mechanism. Gene-editing stocks dropped within hours.

Then came the backlash. Independent biologists called the finding incremental and early. A computational biologist said his team had found the same system years earlier. The preprint had not been peer reviewed, and nobody knows what the system actually does.

What Claude actually found

According to Anthropic's announcement, the lab gave Claude a broad prompt: search a massive DNA database for interesting new examples of reverse transcriptase. Some 950 agents ran for 21 hours, using about 210 million tokens, and surveyed roughly 1.9 billion protein clusters. They pulled out more than 200,000 reverse transcriptases, narrowed those to 3,500 candidates, and produced 20 reports.

One agent then spotted something it could, in Anthropic's telling, see by eye: a long array of evenly spaced, non-coding DNA repeats sitting next to a gene for an odd-looking reverse transcriptase and a partner gene of unknown function. The architecture looks like CRISPR, the bacterial defense system converted into modern gene-editing technology. Early lab experiments showed the repeat array is expressed as short RNAs, hinting it might guide the enzyme the way CRISPR's guide RNAs do. That is, so far, a hypothesis, not a demonstration.

Crucially, Anthropic admits it does not know what ART does. The system's precise function, any biotechnological utility, and even its significance are all still unclear. One hypothesis in the preprint is that ART works like a retron, a known bacterial antiviral system, but the experiments to settle that are still running.

ART is CRISPR-like in its architecture, but there is no evidence that it is CRISPR-like in its function.

Why biologists pushed back

The skepticism arrived fast. Aaron Engelhart, a geneticist at the University of Minnesota, told CNN the results were very interesting but preliminary: they do not demonstrate gene editing or provide a mechanistic picture of what is happening. Philip Kranzusch, a Harvard biochemist, called it a really exciting beginning observation, but not a breakthrough in biology. Steven Salzberg, who heads the Center for Computational Biology at Johns Hopkins, told Bloomberg most biologists do not announce an unproven finding as if it were the next CRISPR.

Lucas Harrington, a gene-editing entrepreneur who co-founded Mammoth Biosciences, argued the hard part of discovery is not spotting a strange gene cluster but figuring out what it does, and said early findings should not carry the framing of a major discovery. Dimitri Perrin, a computer scientist at the Queensland University of Technology, put the technical critique most sharply: the enzyme itself was already known; what is new is recognizing it may form part of a larger system. As he summarized it, ART is CRISPR-like in its architecture, but there is no evidence that it is CRISPR-like in its function.

Even the cheerleaders were measured. Feng Zhang, the MIT and Broad Institute professor who helped pioneer CRISPR editing and reviewed the preprint, called it an exciting example of AI-assisted discovery and said the repeat arrays were genuinely intriguing. (AI is taking over bench work elsewhere too: SLAC is handing its X-ray laser to an AI copilot.) A compliment, not a coronation.

The researcher who says he got there first

The sharpest challenge came from Mario Rodriguez Mestre, a computational biologist at the University of Copenhagen, who told CNN and The New York Times that his team had been studying the same enzymes since 2022, when they turned up in giant viruses that infect bacteria. His group had nicknamed them jumbotrons, and he said its unpublished work describes essentially the same finding.

The twist is that Mestre's team had been using Claude extensively for years, sharing dissertation drafts and manuscripts describing these very systems. His concern, as he put it, is that information from that work somehow reached the models. Anthropic responded that it knows of no previously published work describing ART and that Claude was not trained on user transcripts. Mestre said he was not claiming deliberate theft, but that either his work leaked into training data or considerably more prior knowledge went into the search than the phrase autonomous discovery suggests.

Other biologists echoed the possibility of simple convergence. Seth Shipman, a bioengineer at the Gladstone Institutes and UC San Francisco, said many researchers in the field know about these systems. In fields this crowded, independent rediscovery is routine, and Anthropic's own blog concedes earlier studies had already found the enzyme itself: the claimed novelty is the system-level pattern, the repeat array.

What the episode reveals about AI in science

Read past the dispute, and the experts largely agree on what AI actually did here. Ilya Finkelstein, a molecular biosciences professor at the University of Texas at Austin, told CNN that AI excels at the needle-in-a-haystack search rather than the imaginative breakthrough: agents can pursue a huge number of unproductive paths rapidly, doing the grunt work when orchestrated by experts. Gustavo Sudre, a professor at King's College London, said receiving a broad direction and using its own judgment about what looked interesting is a real shift from the usual analyze this dataset request.

But Sudre added the sentence that should probably be framed: the gap between an interesting pattern and useful knowledge is where most of the work lives. A repeating DNA pattern is a lead, not a result. Confirming function means months of wet-lab experiments, and as Yuval Elani, a biochemical technologies professor at Imperial College London, put it, biology is not maths: intelligence has to be expressed in the physical world through robots, and that automation is not close yet.

Anthropic's own numbers underscore both sides. An AI system surveyed nearly two billion protein clusters in under a day, under the direction of senior molecular biologists, and flagged a pattern experts find worth chasing. But reproducibility is already wobbly: ten reruns of the same search all missed the array, and agents working through tools succeeded as little as a third of the time.

So where does that leave ART? As a research lead, and a test. If follow-up experiments show it can cut or copy DNA in a programmable way, the 21-hour discovery will be remembered as a landmark. If it turns out to be a curiosity, the episode still taught the field something: agents can find the needle, but the haystack is full of needles that look interesting from a distance. The judgment about which ones matter remains, stubbornly, human.

References

CNN, "Anthropic's announcement of a 'Claude-led' CRISPR-like discovery is sparking controversy. Here's why," October 8, 2026 (Anthropic's announcement, expert reactions, and the Mestre dispute). Tapestry News, "Biologist Says Anthropic's AI Enzyme Find Was Already Known," September 28, 2026 (summarizing Anthropic's announcement and The New York Times reporting). Gizmodo, "Claude Found a Mysterious CRISPR-Like System, but Anthropic Can't Say What It's Capable of," September 2026 (the 950-agent search details and outside critiques).