Anthropic says Claude found a CRISPR-like system. Five questions about what is actually proven.
Anthropic's biology team says its agents spotted a new enzyme system in virus DNA. The function is unknown, the enzyme itself was known since 2021, and the work isn't peer reviewed.
Scientific American published a long interview on October 14 with the people behind Anthropic’s “CRISPR-like” enzyme announcement, and it says plainly how early this is. “We have no idea what its function is yet,” said Eric Kauderer-Abrams, Anthropic’s head of life sciences.
That sentence isn’t in the headline version of the story, so here are the questions I’d ask before repeating it.
What did Claude actually do?
Anthropic announced the work on September 23. Researchers gave Claude a prompt to search a huge database of DNA sequences for new reverse transcriptases, enzymes that copy RNA into DNA. About 950 Claude agents worked for 21 hours and used 210 million tokens, according to Anthropic’s post.
One agent noticed a repeating pattern of DNA sitting next to the gene for an odd-looking enzyme. Anthropic’s preprint quotes its note: “I can see by eye a tandem repeat array … that’s a CRISPR-like … repeat array?!” The company calls the whole thing array-associated reverse transcriptases, or ART. It has three parts: the enzyme, a partner gene, and a long array of evenly spaced repeats.
Was it really a discovery?
Partly. Anthropic’s own post says the underlying enzyme, found in a jumbo phage (a very large virus that infects bacteria), “had been identified in previous studies.” Wired reported that microbiologist Jason Gill at Texas A&M and colleagues described the same enzyme in a 2021 paper. What Claude appears to be first at is noticing the repeat array and the accessory protein around it.
I think that’s still interesting. Gill told Wired a person could have done this, but “you have to already have a hypothesis in mind.” Software that reads raw sequence and flags something odd without being told what to look for is a different kind of tool from a search script. It’s also a much smaller claim than “AI discovers new gene editor.”
How close is it to CRISPR?
Not close yet. The preprint says the ART arrays have repeats of 15 to 49 nucleotides and spacers of 120 to 220, while CRISPR spacers are around 30. It also says no cas genes sit near any ART locus. So the resemblance is real but shallow: a repeating layout.
The authors write that they haven’t shown the enzyme is active, or that the short RNAs from the array are its targets. Their agent also flagged the possibility that this is a retron, a different bacterial immune system. Wired noted that retrons are used in some gene editing but aren’t the multi-tool CRISPR is. Feng Zhang, a CRISPR pioneer who read the preprint, called the work “genuinely intriguing” and worth further investigation, which is a polite endorsement of curiosity, not of a tool.
What hasn’t been checked?
Quite a lot. The work is a preprint on bioRxiv, so no journal reviewers have seen it. Only one wet-lab experiment is in it, according to Wired, which showed the array is expressed as separate short RNAs during phage infection. Whether it can edit anything is unknown.
Wired’s subhead quotes an unnamed expert: “The experiments are still in the queue. The PR is already live.” I couldn’t read who said it, since the article is paywalled, but it states the timing problem well. Anthropic’s defense, in the Scientific American interview, is that it published early on purpose so other labs can work out the function.
What does it say about the AI side?
The cost is the part I’d watch. Anthropic’s researcher Nicholas Perry said the tokens and compute were small enough that smaller labs could afford similar runs, and that the same analysis would take a human scientist weeks to months. Those are the company’s numbers, and I haven’t seen anyone outside it reproduce them.
I don’t doubt the search was fast. My doubt is about the next step. Finding an odd pattern took a day, and proving what it does will take years of bench work, as it did for CRISPR itself (Wired puts it at 25 years between the first sequences and a programmable tool). If this turns out to be a useful editor, the story is a good one about speed at the front of the pipeline. If it doesn’t, we’ll have another reminder that announcements arrive before experiments do.
Anthropic also keeps appearing in this site’s safety coverage, for example when its agents did things nobody asked for and the company had to disclose them. That’s the other side of the same capability: agents that notice things on their own.
The next real marker is an experiment showing what ART does, from Anthropic or anyone else. I’d treat everything before that as a promising lead.
Sources
- Claude discovers a novel enzyme system with CRISPR-like repeats, Anthropic, September 23, 2026
- Autonomous AI agents discover reverse transcriptases with tandem repeat arrays, bioRxiv preprint
- Claude just found a CRISPR-like DNA pattern, but even Anthropic’s scientists aren’t sure what it does, Scientific American, October 14, 2026
- Anthropic Says It Discovered a Crispr-Like System. Now What?, Wired, September 29, 2026 (paywalled; I read the opening and the excerpts quoted in search results, not the full article)