Claude noticed a pattern beside a reverse-transcriptase gene that researchers had not previously described as part of a larger system. Human scientists then found evidence that the neighboring DNA array produces distinct short RNAs. It is a credible start to investigating an unfamiliar biological system, not proof that Anthropic has found a new gene-editing tool.
In its September 23 announcement, Anthropic named the candidate system array-associated reverse transcriptases, or ART. It occurs mainly in bacteriophages, viruses that infect bacteria. Its arrangement of repeated DNA resembles CRISPR arrays, but Anthropic says ART’s biological function remains unknown.
The story has two separate results: an AI-assisted search surfaced a promising hypothesis, and early experiments by people support one part of it. Neither establishes what ART does to DNA or whether it could ever be useful for editing genes.
How Claude Narrowed More Than 200,000 Enzymes
Anthropic says it gave Claude a broad assignment: search a large collection of DNA sequences for interesting examples of reverse transcriptases. These enzymes copy RNA into DNA. The scientists did not direct Claude to look for ART or for a CRISPR-like array.
Roughly 950 agents worked for 21 hours, using 210 million tokens, according to Anthropic. They surveyed more than 200,000 reverse transcriptases, identified approximately 3,500 candidate systems, and narrowed those to 20 candidates for detailed, human-readable reports. Those figures describe successive stages of the search, not 200,000 laboratory experiments.
The agents examined sequence data, investigated enzyme families, checked nearby genes and consulted existing research. The team used Claude Science and Claude Code, sometimes with its own software to coordinate many Claude sessions. Human scientists supplied the initial research direction and reviewed the resulting hypotheses.
One candidate drew attention when an agent examined raw DNA near an unusual reverse-transcriptase gene. It noticed regularly repeated sequences in the neighboring DNA, then counted the repeats, examined their spacing, compared the arrangement with known systems and searched for earlier descriptions before submitting its report.
The agent flagged a pattern in data and pursued it as a possible clue; it did not simply retrieve a named system from a paper. The 21-hour figure covers the agent search, not the time required for human assessment or laboratory follow-up. Anthropic has not provided a like-for-like comparison showing what the same investigation would have cost or how long it would have taken by other methods.
