Anthropic’s AI Biology Claim Meets the Lab-Test Question
Anthropic says a large group of Claude agents found an unusual biological system in genomic data. The finding has attracted attention because its pattern resembles part of the machinery behind Crispr, but the distance between spotting a pattern and demonstrating a useful gene-editing tool remains large. Scientists interviewed by WIRED see a potentially interesting result and a strikingly fast search. They do not agree that Anthropic has established a new Crispr system, and the essential experiments have not yet been done.
What the AI agents actually found
Anthropic announced the result on September 23. According to the company, roughly 950 Claude agents worked in parallel for 21.5 hours, searching extensive genomic databases for examples of reverse transcriptases. These proteins copy RNA into DNA. The agents initially surfaced more than 200,000 possibilities, then narrowed the field through several rounds of selection to an unusual family associated with long, repeated DNA sequences.
Anthropic calls the candidate ART, for array-associated reverse transcriptases. It appears in jumbo phages, large viruses that infect bacteria. The repeated sequences drew comparisons with Crispr, but resemblance is not a demonstration of equivalent function. The company itself says it does not yet understand what the system does. Its technical report includes one physical experiment and has not been peer-reviewed, WIRED reports.
The search also needs careful credit. A 2021 paper by microbiologist Jason Gill and colleagues had identified the same reverse transcriptase. What may be new in Anthropic’s work is the identification of nearby repeats and the hypothesis that they form a broader system. That distinction matters: the agents did not necessarily discover every component from a blank slate, nor did they show that ART cuts or edits DNA.
Why scientists are reserving judgment
Researchers who spoke to WIRED give the work different weights. Fyodor Urnov of the University of California, Berkeley praised Anthropic for publicizing the candidate. Stanford professor Le Cong emphasized that the laboratory tests still lie ahead. Seth Shipman of the Gladstone Institutes called the method of finding the system more novel than the biological object itself. Gill said the proposed relationship to known Crispr systems is not obvious and that Anthropic must establish any editing activity experimentally.
One alternative discussed in the report is that the pattern relates to a retron. Crispr and retrons are different bacterial immune systems, even though both can interest gene-editing researchers. The label attached to a promising sequence can therefore run ahead of its biological characterization. If ART eventually proves useful, scientists would still need to determine its mechanism, whether it can be controlled, and how reliably it acts. None of those outcomes follows from the computational search alone.
The reported speed is nevertheless notable. Shipman said a manual search for unfamiliar reverse transcriptases could take months. Claude’s pattern-finding approach appears to have compressed that discovery stage into about a day. But the work was designed and interpreted by human scientists. Treating the process as a wholly autonomous discovery would obscure their choices about the question, the data, and the meaning of the result.
A promising lead with unresolved evidence
The claim also raises a reproducibility question. WIRED notes that outside researchers cannot evaluate exactly what material the model was trained on. That makes it harder to separate a genuinely new inference from a connection that may already have appeared in the information the system encountered. Anthropic told another publication, as recounted by WIRED, that Claude was not trained on user transcripts and that its molecular biology team had no access to them. The broader limits on independent inspection remain.
Cong suggested that a more demanding test of an AI scientist would ask whether a model trained without later discoveries could still find something useful. In this case, the immediate next step is less philosophical: laboratory work must establish what ART does. Anthropic has offered a candidate and a hypothesis, not a validated editing platform. The difference will be decided by experiments that can be examined and repeated, not by how convincingly a sequence resembles a familiar biological tool.
Anthropic’s chief executive has speculated that Claude might someday control lab equipment to test hypotheses itself. WIRED presents that as a distant possibility, not a capability demonstrated by this finding. For now, the interesting result is a fast, AI-assisted search that produced a testable biological lead. The claims about Crispr-like function and practical value remain open.
