Anthropic’s Claude finds a CRISPR-like system, not a gene-editing tool

A pattern, not yet a tool

Anthropic asked Claude to search genomic databases for new examples of reverse transcriptases, proteins that copy RNA into DNA. The agents first identified more than 200,000 candidates, then narrowed the search to a family with a long stretch of repeating DNA sequences resembling CRISPR. Anthropic named it ART, for arrayed reverse transcriptases. The system occurs in giant phages, viruses that infect bacteria.

One agent noted that the repeats looked like CRISPR and suggested the system might be a retron, a type of bacterial immune system. Retron-based gene-editing methods already exist, though retrons are not as versatile as CRISPR.

The key caveat is straightforward: Anthropic does not yet know what ART does. Its technical report includes one laboratory experiment and has not been peer-reviewed. Whether the system can edit genes, or would be useful if it can, remains untested.

The finding also has a history. Microbiologist Jason Gill of Texas A&M University and colleagues described the same reverse transcriptase in a 2021 study of giant phages. Claude may have added value by spotting nearby repeats and proposing that the enzyme could belong to a CRISPR-like system. That is a meaningful research lead, but not the same as discovering the enzyme itself.

What the AI did—and what it didn’t

Scientists quoted in the report draw a distinction between finding a pattern and establishing what it means. Seth Shipman, an associate investigator at the Gladstone Institutes, called the result interesting but said the novelty was in the search method, not the system. His lab has used retrons to build gene-editing systems, showing that this broader direction is plausible.

Gill says models can spot patterns in data better than a person simply looking at it. But he also says ART has no obvious connection to known CRISPR systems, and Anthropic still has to demonstrate that it edits genes. Leon Kong, a Stanford professor who studies AI in genetic engineering, made a similar distinction: the experiments are still pending while the publicity has begun.

The discovery is Anthropic’s first from a research group established earlier this year. The company has opened a lab focused on drug development and says work on ART will continue. But the announcement leaves a basic question unanswered: how much of the result came from Claude’s pattern recognition, and how much from the scientists who chose the search, interpreted the output and tested the candidate?

Reproducibility is another gap. Researchers cannot assess the data used to train Claude, a limitation that matters when scientific journals increasingly ask authors to publish code so others can reproduce their work. One researcher who had studied the same enzymes suspected the model might have learned from unpublished results he had shared with a public version of Claude. Anthropic told The New York Times that it did not train Claude on user conversations and that its molecular biology team cannot access them.

The long road from discovery to use

CRISPR itself is a reminder of how much work follows an initial biological observation. Researchers identified the first CRISPR sequences in bacteria in 1987. In 2012, Jennifer Doudna and Emmanuelle Charpentier showed that CRISPR could be turned into a programmable tool for cutting DNA. It is now widely used in labs and being tested in dozens of clinical trials, but only one drug based on the technology had reached the market, approved in late 2023.

Anthropic CEO Dario Amodei has suggested that Claude might eventually conduct experiments by operating lab equipment on its own. He said that prospect was still a long way off, and Anthropic said its labs operate at minimal biosafety levels, with nothing that poses a serious danger to people.

I think the more consequential test is not whether Claude can find another promising pattern, but whether it can find one without being told what kind of thing to look for. Kong’s proposed test is to train a model only on knowledge available before CRISPR was discovered, then see whether it can identify CRISPR independently. Until a system can pass that kind of test—and its predictions survive laboratory work—AI’s role in biology looks more like a faster way to generate hypotheses than an autonomous engine of discovery.