Claude’s CRISPR-like enzyme discovery: what was found, and what is still unknown

The striking result is an AI-led path from a sequence anomaly to a wet-lab experiment. It is also an early research report, not a new gene-editing treatment or a fully autonomous laboratory.

Anthropic announced on September 23 that Claude had helped identify a previously uncharacterized enzyme system associated with repeated DNA sequences. The company called it array-associated reverse transcriptases (ART) and said its layout resembles a CRISPR array. The announcement became a major AI-science story because it describes a concrete discovery path: a model searched sequence data, noticed an unusual pattern, proposed a candidate, and human scientists tested it in a laboratory.

What exactly did Claude find?

The system is based on a reverse transcriptase, an enzyme that copies RNA into DNA. Anthropic says the underlying reverse transcriptase in a jumbo bacteriophage had been identified before, but the surrounding arrangement had not been characterized as a system.

The new part is the combination of three elements:

• a reverse transcriptase;
• a nearby partner gene of unknown function; and
• a long, organized array of repeated non-coding DNA sequences.

Anthropic’s first experiments showed that the repeat array is expressed as distinct short RNAs. That is suggestive, but the company says further experiments are needed to determine what ART actually does.

How the discovery was made

According to Anthropic, researchers gave Claude a high-level prompt to search a large DNA database for unusual reverse transcriptase systems. Roughly 950 agents then worked through the data for 21 hours, consuming 210 million tokens.

The broader funnel was substantial. Claude agents gathered more than 200,000 reverse transcriptases, selected 3,500 candidate systems, and reduced them to 20 compelling candidates with written reports. One agent noticed a tandem repeat array beside an unusual reverse transcriptase. The model compared the spacing and layout with known systems, searched the literature, and wrote a report for human review.

That sequence matters. It was not a prompt for the answer. It was a large, auditable search process followed by a proposed anomaly, with human scientists deciding what deserved physical testing.

Humans still ran the laboratory

This is the part most likely to disappear from viral summaries. Anthropic says its Bay Area lab works at biosafety levels BSL-1 and BSL-2, does not handle human-infectious pathogens, and performs all laboratory work through human scientists. The models helped interpret data and generate hypotheses, but they did not autonomously operate the equipment.

TechCrunch highlighted the same boundary. It reported that the lab did not let Claude run loose in the physical workspace and noted Anthropic’s statement that autonomous equipment control may be possible in the future with appropriate safeguards, but is not happening today.

So “AI discovered an enzyme system” is a compressed description, not a full account of the workflow. A more accurate shorthand is: AI narrowed a vast search space and generated a testable lead; humans characterized the biological system.

Why it matters for AI agents

The case is more useful as an example of agentic scientific work than as proof that AI has become an independent scientist. The workflow combines several capabilities:

1. Large-scale retrieval: agents survey sequence families and gather candidates.
2. Pattern detection: one agent flags an unusual arrangement that does not fit the expected format.
3. Critical filtering: additional agents eliminate weak candidates and prepare readable reports.
4. Human scientific judgment: researchers select candidates, run experiments and revise the hypothesis.

Each stage is measurable. Teams can count candidates reviewed, time spent, reasons for rejection, experiment outcomes and false positives. That is the kind of structure serious AI-for-science programs need: not just a dramatic final result, but a workflow whose intermediate decisions can be checked.

What is confirmed—and what is not

Reported by Anthropic and supported by its preprint: the sequence organization, the first expression results, the size of the computational search and the human-led laboratory process.

Still unknown: ART’s primary biological function, whether it performs useful programmable operations, whether it can become a biotechnology tool, and how its importance compares with other repeat-associated systems.

Anthropic’s technical report is the correct starting point for researchers. The company says the system is found mainly in bacteriophages—the viruses that infect bacteria—and that further experiments are underway. The preprint should be evaluated as a preprint, not a settled discovery with immediate applications.

The skeptical reading is also important

Kevin Blake, a microbiologist at Washington University School of Medicine, told Al Jazeera that “CRISPR-like” does not imply “next CRISPR.” He noted that many CRISPR-like natural sequences remain poorly catalogued and that the current evidence does not indicate a rival editing technology or practical therapeutic application.

Stanley Qi of Stanford described the process as exciting because AI recognized a complex pattern and pursued it comprehensively. Both reactions can be true: the workflow demonstrates a valuable capability while the biological result still requires independent scrutiny.

What readers should watch next

Four developments would make this story more consequential:

• Independent replication: other labs should confirm the sequence association and expression results.
• Functional experiments: researchers need to determine what the enzyme system does in bacteriophages.
• Comparison with prior work: Stanford researchers have described related repeat-associated systems, so novelty should be assessed against the wider literature.
• Open methods: the preprint, prompts, candidate-selection process and experimental protocols will be more informative than a launch-video claim.

Bottom line: Anthropic’s report is a credible signal that biological agents can compress parts of discovery work. It is not evidence that Claude independently conducts science or that ART is “the next CRISPR.” The durable story is the loop: AI-generated hypotheses, traceable filtering and human experiments.

Prices and benchmark figures are standard published rates as of September 25, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.