Anthropic CRISPR-Like Discovery Claim Meets Scientist Skepticism

What Anthropic announced

Anthropic has made its first major bid to show that large language models can do for biology what they have done for coding. On September 23, the company said its Claude model had identified an enzyme system in viral DNA with properties it described as "reminiscent of Crispr," the Nobel Prize-winning gene-editing technology 2. The detected signature resembles the microbial machinery behind tools such as CRISPR-Cas9, which scientists use to modify the DNA of living organisms 1.

The scale of the search is the headline number. According to Anthropic, about 950 Claude agents ran in parallel and surfaced the sequences of interest in roughly 21.5 hours 2. Nature's news coverage described the effort as more than 21 hours of AI agents combing through DNA sequence databases 3. Chief executive Dario Amodei framed the result ambitiously on X, calling it "the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism" 1.

This is the first finding from a research group Anthropic set up earlier this year 2. That context matters. It explains why the company is promoting the result so loudly, and why researchers are examining it closely.

Why scientists are pushing back

The pushback concerns how much the result proves and how it was presented. CNN reports that experts not involved in the work called the finding early and incremental, "far from a genuine breakthrough" 1. Wired captured the criticism with one expert's line: "The experiments are still in the queue. The PR is already live" 2.

As described, the work is computational. The agents flagged sequences that look like a CRISPR-type system. Showing that the enzyme actually cuts or edits DNA, and does so usefully and precisely, requires laboratory experiments that the critics say have not yet been completed 2. In molecular biology, many sequences that look promising on a computer turn out to have no useful function once tested. The term "CRISPR-like" draws on the reputation of one of the most celebrated tools in modern biology, while the evidence so far supports a much narrower claim about sequence similarity.

The response has not been entirely hostile. Fyodor Urnov, a gene-editing specialist at UC Berkeley and director of therapeutic R&D at the Innovative Genomics Institute, said he "sincerely" complimented Anthropic for sharing the discovery 2. Wired notes, however, that IGI collaborates with Anthropic, although it was not involved in this specific finding 2. Readers should weigh that relationship alongside his endorsement.

The main disagreement among observers is about framing, not facts. Nobody quoted disputes that the agents found something interesting. The argument is over whether a corporate announcement and a CEO's social media post should come before the experimental validation that would normally accompany a claim this large.

The CRISPR bar is high

The bar for anything called "CRISPR-like" is high. Current research shows that gene editing is still being refined after more than a decade of intensive work. Recent studies in Nature Portfolio journals report that CRISPR-Cas9 off-target edits and chromosomal translocations vary widely between individual cells and organs, which raises ongoing safety concerns for therapies 3. Other groups are building new base editors to widen the range of DNA changes that are possible 3.

In that field, a newly flagged enzyme system is a starting point. Even well-characterized editors require years of work on specificity, delivery and safety before they matter clinically. A candidate found in silico has not yet entered that process.

Our read

The announcement is a real but modest milestone, and the marketing has outpaced it. Running hundreds of agents in parallel to sift genomic databases in under a day is a credible demonstration of how AI could speed up the hypothesis-generation stage of biology, and that may be the most durable part of the result. Calling it a potential new gene-editing mechanism goes beyond what has been shown.

AI companies are increasingly arguing that their models can transform the life sciences, as CNN notes 1. The strongest way to support that argument is with validated wet-lab results, ideally peer-reviewed. Announcing results before the experiments are finished risks the credibility these companies need with the scientific community.

If the lab experiments confirm meaningful editing activity, Anthropic will have a strong case. If they do not, this episode will be a reminder that biology is slower to deliver results than software.