Claude Found the ART System: What Was Discovered in Bacteriophage DNA

On September 23, 2026, Anthropic announced that Claude had found a previously undescribed enzyme system in bacteriophage DNA. It was named ART (array-associated reverse transcriptases). The system’s architecture resembles CRISPR, but what it does remains unknown. Below is an account of how the search was conducted, what was confirmed in the laboratory, and where the limits of this result lie.

Where the task came from

In spring 2026, Anthropic assembled a research group and built its own “wet” laboratory in the San Francisco Bay Area. They tested a simple hypothesis: can an AI agent accelerate the kind of discovery in biology that usually begins with noticing something unusual in the data?

This approach has a long history. Restriction enzymes were first discovered as part of bacterial defense, and later began to be used to cut DNA. Taq polymerase was isolated from a Yellowstone hot-spring bacterium, and PCR was built on it. CRISPR itself was first noticed as a set of unexplained repeats in a bacterial genome.

One of the first projects involved searching for unusual reverse transcriptases (RTs)—enzymes that copy information from RNA back into DNA. Claude was given a broad task: study a huge sequence database and find interesting, poorly studied variants. No one specifically requested a CRISPR analog.

How the search proceeded

The agents then worked with almost no human intervention. They compared RT families, examined neighboring genes, and selected the most unusual variants.

According to the company’s own estimate, this analysis would have taken a human specialist anywhere from several weeks to several months.

The moment of discovery

One of the agents noticed an unusual RT and decided to examine the DNA next to it. There it found a long pattern of repeating regions, resembling a CRISPR array. The agent counted the repeats, measured the distances between them, compared the finding with known systems, and checked the scientific literature to see whether anyone had described it before. Finding no matches, it prepared a report for the scientists.

What ART consists of

The system is found mainly in bacteriophages, or viruses that infect bacteria. It has three components:

• a reverse transcriptase enzyme;
• a neighboring partner gene whose function is unknown;
• a long array of evenly spaced DNA repeats.

The array attracted the researchers’ greatest interest. In CRISPR, a similar array stores a set of RNA guides that allow the system to be programmed to target a specific section of the genome. Initial experiments showed that the ART array is also transcribed into a set of individual short RNAs. This suggests a programmable mechanism, but is not proof.

The reverse transcriptase itself had been encountered before, in a large bacteriophage. According to Anthropic, the novel finding was the combination: no one had linked this RT to an array of noncoding repeats and a partner protein.

What was confirmed in the laboratory

Laboratory testing showed that this was a real, previously undescribed system, rather than an artifact of the analysis. The proteins were expressed in standard laboratory strains, followed by biochemical and structural characterization. The results were interpreted by Claude again.

The preprint was published simultaneously with the announcement and has not yet undergone peer review. Feng Zhang, a professor at MIT and the Broad Institute and one of the pioneers of CRISPR-based genome editing, read it. He called the work an interesting example of AI agents participating in biological discoveries and noted that the identified arrays of repeated RNAs warrant further study.

How the division of labor works

The laboratory operates only at BSL-1 and BSL-2 biosafety levels and does not work with pathogens dangerous to humans. People conduct all physical experiments. Claude handles data analysis and hypothesis generation, and its workflow follows the same pattern from project to project:

1. reads the literature on the selected protein family and reproduces known results;
2. searches for undescribed members of the family and unusual genomic neighbors;
3. writes a report for each interesting candidate describing its presumed function and the supporting evidence;
4. critically tests its own hypotheses and discards most of them;
5. passes only the most promising candidates on to people for experiments.

The fourth step appears to be the key one. Of the 3500 potentially new systems, only one reached the laboratory: without rigorous self-filtering, the scientists would have been overwhelmed by plausible but empty candidates.

What the result does not prove

• Anthropic is not claiming that ART will become a new gene-editing tool.
• The system’s function and role in the bacteriophage’s life are unknown. The team is currently trying to connect it to short RNAs.
• The preprint has not been peer-reviewed, and the conclusions may be refined.

What the experiment showed conclusively: the agent can independently work through a vast body of genetic data, notice a pattern it was not asked to look for, check it against the literature, and produce a candidate that withstands testing in a tube.

What this means for research teams

• Set the task more broadly than seems necessary. ART was found while searching for “interesting RTs,” not for a CRISPR analog. A narrowly defined task would have ruled out the discovery.
• Build self-criticism into the pipeline. The agent should reject its own hypotheses before they reach people. Otherwise, cheap idea generation runs into expensive validation.
• Require a report with evidence. For each candidate: its presumed function, what it is based on, and what has been checked against the literature.
• Leave experiments to people. The model speeds up exploration and selection, but confirmation still comes only from experiments.