Anthropic says Claude found a Crispr-like system; scientists urge caution

In a September 23 announcement, Anthropic claimed that its large language model Claude had identified an enzyme system with properties "reminiscent of Crispr," the Nobel Prize-winning gene-editing tool. According to Anthropic, about 950 Claude agents running simultaneously found the interesting genetic sequences in 21.5 hours. It is the first finding to come out of a research group Anthropic formed earlier this year. The company has also set up a wet lab for drug discovery.

The process, as Wired describes it: researchers at Anthropic prompted Claude to search huge genomic databases for "interesting new examples" of reverse transcriptases, proteins that copy RNA into DNA. The agents initially identified more than 200,000 possible reverse transcriptases, then picked out several thousand that appeared to be new. They narrowed the set further and found an "unusual" reverse transcriptase family with a long region of repeat DNA sequences resembling Crispr. Anthropic calls the system ART, short for array-associated reverse transcriptases. It is found in jumbo phages, large viruses that infect bacteria. In Anthropic's material, an AI agent wrote: "I can see by eye a tandem repeat array ... that's a Crispr-like ... repeat array?!" The agent also acknowledged that the system could be a retron, as the technical report notes. Crispr and retrons are both bacterial immune systems, and retrons have some use in gene editing, but they are not the same multi-tool Crispr is.

Anthropic itself says the work is not finished. In an X post it said: "We don't yet understand what this system does," adding that only a handful of known systems share its features, and all of them can cut, copy and paste DNA. One physical experiment was included in a technical report, which has not been peer-reviewed. It remains unknown whether the system can be used as a gene-editing tool and, if so, whether it is a useful one.

Outside scientists were mixed. Fyodor Urnov of the University of California, Berkeley, and its Innovative Genomics Institute (IGI), said: "I sincerely compliment Anthropic for telling the world about their discovery." IGI collaborates with Anthropic but was not involved in this work. Le Cong, a Stanford professor focused on integrating AI into genome engineering research, was sharper: "The experiments are still in the queue. The PR is already live." He compared it to scanning all the sand on Santa Monica Beach for a diamond: AI found something shiny, and only lab work can show whether it is glass or a diamond.

Seth Shipman, associate investigator at the Gladstone Institutes, does not think Anthropic found a new Crispr system, but calls the discovery interesting: "The novel thing is how they found it, not what it is." His lab has used retrons to make gene-editing systems, so using them that way is possible. Finding new reverse transcriptases by mining genome databases manually can take months, so doing it in a day is impressive, he says. He cautions against crediting Claude alone: "I think we have to be careful about saying that Claude autonomously discovered something, because there are scientists involved in the study."

Jason Gill, a microbiologist at Texas A&M University, and colleagues had already identified this same reverse transcriptase in a 2021 paper on jumbo phages. What may be new is that Claude spotted repeats around it hinting at a Crispr-like system. Gill says: "These models are good at finding patterns, better than a person staring at it with their eyeballs can. 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 it is on Anthropic to prove it has gene-editing activity. Cong credits the scientists for picking a problem suited to Claude's pattern-seeking. Wired notes that the model could work on similar large-data projects, but that does not mean Claude will suddenly speed up discoveries across the life sciences.

Wired also raises transparency. The scientific community cannot evaluate what the model was trained on, which makes the findings unusually opaque at a time when major journals require scientists to publish their code. One 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 told the New York Times that Claude was "not trained on any user transcripts, and our molecular biology team has no such access, either."

Anthropic CEO Dario Amodei wrote on X: "Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment." Wired calls that a fairly distant prospect. Amodei quickly followed up by noting that Anthropic's labs are at the lowest biosafety levels. Wired closes with history: it took 25 years from the discovery of Crispr sequences in bacteria in 1987 until Jennifer Doudna and Emmanuelle Charpentier showed in 2012 it could be a programmable DNA-cutting tool, and only one Crispr drug is on the market, approved in late 2023. Cong's suggested better test of an AI scientist: train a model only on knowledge from before Crispr was discovered and see whether it finds Crispr.

Key facts

Anthropic says about 950 Claude agents running simultaneously found the ART system (array-associated reverse transcriptases) in 21.5 hours; it appears in jumbo phages.
The agents first flagged more than 200,000 possible reverse transcriptases, then several thousand that looked new, before landing on one family with Crispr-like repeats.
Only one physical experiment is in Anthropic's technical report, which is not peer-reviewed; whether ART works as a gene-editing tool is unknown.
Shipman does not think it is a new Crispr system, and Gill says the same reverse transcriptase appeared in a 2021 paper he co-wrote on jumbo phages.
Le Cong: "The experiments are still in the queue. The PR is already live."

Why it matters

Tech executives have long promised that AI will speed up biology, and this is a concrete claim from a frontier lab. Anthropic says a swarm of agents did in 21.5 hours what Shipman says can take months by hand. The scientists Wired spoke to see the method as the interesting part, not the enzyme: in Shipman's words, the novel thing is how it was found, not what it is. The open question is whether an AI can find something useful without knowing exactly what it is looking for.

Who it affects

Gene-editing and microbiology researchers, who are being asked to judge a claim before the lab data exists. Anthropic's new research group and its wet lab for drug discovery, since this is that group's first finding. Also the wider scientific community, which cannot inspect what the model was trained on.

How to use it

There is nothing to use yet. Whether ART can serve as a gene-editing tool, or a useful one, is unknown. The practical takeaway is the method: prompt a model to mine large genomic databases for a defined class of proteins, then take the candidates to the lab. Cong says the model could handle similar large-data projects, though that does not mean broad acceleration across the life sciences. No price or access terms are given.

How solid is it

Weak so far. The claim is Anthropic's own, made in an announcement and a technical report that has not been peer-reviewed and includes one physical experiment. Anthropic itself says it does not yet understand what the system does. Shipman does not think it is a new Crispr system, and Gill says it appears to have no obvious relationship to any known Crispr system and that Anthropic must prove gene-editing activity. The same reverse transcriptase was described in a 2021 paper, so what is new is the repeats Claude noticed around it. Urnov praised Anthropic for disclosing the discovery.

Risks and caveats

The Crispr comparison may oversell it: the agent itself noted the system could be a retron, and Wired says it is not surprising the blog post played up the Crispr angle. Shipman warns against saying Claude discovered it autonomously, since scientists were involved. Training data cannot be evaluated externally, and one researcher suspects his unpublished findings shared with the public Claude were used; Anthropic denies training on user transcripts. Amodei's idea of Claude running lab equipment raises safety and oversight questions. Translating a discovery into medicine is slow: 25 years from Crispr's discovery to a programmable tool, and one approved drug so far.

“The experiments are still in the queue. The PR is already live.”

— Le Cong, professor at Stanford University