Anthropic's Claude Found a Crispr-Like System in 21.5 Hours. Scientists Are Split

Anthropic says about 950 Claude agents running simultaneously found an enzyme system with properties "reminiscent of Crispr" in 21.5 hours. The September 23 announcement is the first discovery from a research group Anthropic formed earlier this year, and it lands at the center of a debate the field has been waiting to have: can AI models actually accelerate biology, or just accelerate announcements about biology?

The claim has teeth, but they are small ones so far. According to Anthropic, Claude agents initially identified more than 200,000 possible reverse transcriptases — proteins that copy RNA into DNA, the reverse of the normal cellular process — after researchers prompted the model to search huge genomic databases for "interesting new examples." The agents narrowed those down to several thousand that appeared new, and eventually landed on an "unusual" reverse transcriptase family containing a long region of repeat DNA sequences that resemble Crispr. Anthropic calls the system ART, short for array-associated reverse transcriptases. It lives in jumbo phages, large viruses that infect bacteria.

One of the agents appeared to register the significance of the find itself. "I can see by eye a tandem repeat array … that's a Crispr-like … repeat array?!" an AI agent wrote, according to the technical report. The same agent acknowledged the system could be a retron — a different class of bacterial immune system. Crispr and retrons are both bacterial defenses, and while retrons have some utility in gene editing, they are not the multi-tool Crispr turned out to be. That distinction did not stop Anthropic's blog post from leading with the Crispr comparison.

The PR arrived before the experiments

The discovery divides researchers less on whether it is interesting than on whether it is proven. "The experiments are still in the queue. The PR is already live," says Le Cong, a professor at Stanford University focused on integrating AI into genome engineering research.

Anthropic has been candid about the preliminary nature of the work. "We don't yet understand what this system does," the company wrote on X, "but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA." The technical report — which has not been peer-reviewed — includes a single physical experiment, conducted at the wet lab Anthropic has built for drug discovery. Nobody yet knows whether ART works as a gene-editing tool, or whether, if it does, it is a useful one.

"Let's say we are on Santa Monica Beach and trying to scan through all the sand to find a diamond," Cong says. "AI found this thing that looks very shiny, and then you have to go back to the lab to know—is it glass? Is it a diamond?"

Fyodor Urnov, a gene-editing expert at the University of California, Berkeley and director for therapeutic R&D at its Innovative Genomics Institute, is more generous, at least about the disclosure. "I sincerely compliment Anthropic for telling the world about their discovery," he says. IGI collaborates with Anthropic but was not involved in the new finding.

Not a new Crispr — and not entirely new

Seth Shipman, associate investigator at the Gladstone Institutes, does not believe Anthropic found a new Crispr system. He still finds the discovery worthwhile — for its method. "The novel thing is how they found it, not what it is," he says. His lab has used retrons to build gene-editing systems, so the retron interpretation does not diminish the finding's potential utility.

The speed matters. Identifying new reverse transcriptases can take months of manual genome-database mining, Shipman says, so compressing that to roughly a day is a genuine result. But he cautions against crediting the model alone. "I think we have to be careful about saying that Claude autonomously discovered something, because there are scientists involved in the study," he says.

That caution has a concrete basis. The reverse transcriptase at the center of ART was previously identified by Jason Gill, a microbiologist at Texas A&M University, and his colleagues in a 2021 paper on jumbo phages. What may be genuinely new is that Claude spotted the repeats around the enzyme that suggested membership in a Crispr-like system.

"These models are good at finding patterns, better than a person staring at it with their eyeballs can," Gill says. "A person could do all this but you have to already have a hypothesis in mind." He adds that ART appears to have no "obvious relationship" to any known Crispr system, and that the burden now sits with Anthropic to prove the enzyme actually has gene-editing activity.

Opacity, training data, and a suspicious researcher

The finding also exposes an unusual opacity problem. The scientific community cannot evaluate what information Claude was trained on, at a time when major journals require scientists to publish their code freely in the name of reproducibility. That opacity has already produced friction: a researcher who had been studying these exact enzymes became suspicious that the model was trained on his work, because he had shared unpublished findings with the public version of Claude. Anthropic denied this to the New York Times: "Claude was not trained on any user transcripts, and our molecular biology team has no such access, either."

Cong credits Anthropic's scientists with picking the right kind of problem for Claude — a large-scale pattern search that matches the model's strengths. That does not mean Claude will suddenly accelerate discoveries across the life sciences.

The long arc from sequence to medicine

Even a verified discovery would begin a long road. It took 25 years from the initial identification of Crispr sequences in bacteria in 1987 until Jennifer Doudna and Emmanuelle Charpentier demonstrated in 2012 that Crispr could work as a programmable tool for cutting DNA. Crispr is now ubiquitous in laboratories and in dozens of clinical trials targeting cardiovascular conditions, cancers, autoimmune diseases, and rare disorders. Exactly one Crispr-based drug has reached the market, approved in late 2023.

Anthropic CEO Dario Amodei has already sketched a more ambitious endpoint. "Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment," he wrote on X. He quickly noted that Anthropic's labs operate at the lowest biosafety levels and contain nothing greatly harmful to humans. Even Anthropic's own framing places that scenario in a fairly distant future.

The deeper open question is whether any AI model can find something useful without knowing what it is looking for. What made Crispr revolutionary was the identification of a previously unknown biological system. Claude found a pattern near a known enzyme.

"If you have an AI that only trained on knowledge from before people ever discovered Crispr, and then that AI actually discovered Crispr, that seems to be a better setup to test an AI scientist," Cong says. Until a model passes that kind of test — and until ART shows real gene-editing activity in a lab — Anthropic's discovery stands as an impressive demonstration of AI-driven search, with the diamond-or-glass question still unanswered.