Anthropic's Claude Found a Novel Enzyme in 21 Hours. Scientists Are Split

Anthropic says 950 Claude AI agents discovered a previously uncharacterized CRISPR-like enzyme system in roughly 21 hours, after scanning about 200,000 reverse transcriptase samples from public databases. The company named the system "array-associated reverse transcriptases," or ART, and human scientists confirmed it as a novel find. But researchers who spoke to CNET say the discovery's significance is far from settled — and one team claims it found the same enzyme in 2022.

The stakes are considerable. CRISPR is one of the biggest scientific discoveries in modern history; it took nearly 40 years from the system's identification in 1987 to the first approved CRISPR medicine, in the UK in 2023 and then from the US Food and Drug Administration in 2024. If AI can compress the front end of that pipeline — finding candidate systems in days instead of months — the economics of biological discovery change.

What did Claude actually find?

Anthropic says human involvement was limited to the prompts and the lab work. The agents themselves scoured publicly available databases to identify something unique. Spotting the enzyme took about 21 hours. Human scientists then confirmed it had never been observed by science before.

The enzyme was found in a bacteriophage — a virus that attacks bacteria — specifically one with a large DNA genome, commonly known as a jumbo phage. To understand why that matters, some background is needed.

CRISPR, short for clustered regularly interspaced short palindromic repeats, is a bacterial defense system. When a bacterium survives a viral infection, it copies a snippet of the virus's DNA into its own genome, building a library across generations. When a virus attacks again, the bacterium deploys an RNA carrying the matching snippet, which guides a cutting protein to slice the viral DNA and stop replication.

Reverse transcriptase is different. It is an enzyme that uses RNA as a template to build DNA — the reverse of the normal direction of information flow, where DNA is transcribed into RNA. Reverse transcriptase is a defining feature of many viruses, including HIV; understanding how HIV uses it to replicate led to some of the first treatments and still underpins modern therapies.

The connection: CRISPR contains many repeating DNA fragments, which is how scientists identified it in the first place. In the decades since, they have only found CRISPR-like arrangements in bacteria and archaea. What Claude noticed was a reverse transcriptase sitting next to a long array of repeating DNA — a layout strikingly similar to CRISPR.

"What Claude noticed is another type of protein, a reverse transcriptase, sitting next to a long array of repeating DNA, which is very similar to how CRISPR is laid out," a representative for Profluent, an AI lab that made OpenCRISPR-1, the first AI-generated gene editor, told CNET in an email. "This particular reverse transcriptase had been seen before. The array next to it and a partner protein alongside it are what Anthropic is sharing."

How big of a deal is it?

Feng Zhang, professor at MIT and the Broad Institute, offered a measured endorsement: "This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation."

The problem is twofold. First, timing. Scientists took roughly 20 years to figure out how CRISPR worked, then nearly another two decades to build the first medicine from it. The same lag could apply to ART.

"Scientists are finding interesting sequence patterns all the time, relatively speaking," Bonnie Berger, head of the Computation and Biology group at MIT's Computer Science and AI Lab, told CNET in an email. "The bottleneck is then figuring out the purpose and function of these systems using deeper analysis, expert biological knowledge, and complementary targeted experiments."

Anthropic acknowledged the gap. "Our work to understand the primary function of ARTs is ongoing," the company said in a statement. "However, we think it is important to share such findings early, both to demonstrate Claude's capabilities and to give the broader community insight into what we're working on."

Was it really a first?

The second problem is priority. Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, told The New York Times that his team found this enzyme in 2022.

"They are the same systems we have been studying for years," Mestre told the Times. He points to a patent containing the research, on which he is listed as a co-inventor.

Anthropic says it was unaware of that research.

Mestre's team raises a sharper question: whether Claude found the enzyme on its own, or whether the model had been trained on the unpublished research the Copenhagen scientists had uploaded to Claude for years to assist their own work — data that, in Mestre's framing, would have sent Claude to the enzyme like a homing beacon. "I mean, coincidences happen," Mestre told the Times. His team is shutting down its Claude projects and moving to other models, just in case.

The pickings for alternatives are slim. OpenAI was accused of something similar while solving the Navier-Stokes Millennium Problem, and could not rule out that its model had been trained on "deidentified data derived" from Levent Alpöge and Tristan Buckmaster, the mathematicians who actually solved it.

Where the experts do agree

Every expert CNET spoke to agreed on one point: the speed is real. Anthropic said the same search could have taken human researchers weeks or months. Berger put the scale of the search space in context: "We have hardly scratched the surface of exploring the biological systems represented in species sequenced outside of a handful of model organisms. So, many accomplished scientists could spend their careers studying interesting sequence patterns without getting to this one."

Berger also flagged a risk for the field. Lazy AI research, she warned, can lead to "an overwhelming volume of low-impact observational findings that smothers important, targeted research on promising disease treatments."

That tension — high-throughput AI discovery versus slow, targeted validation — will define the next phase of computational biology. Anthropic will have to wait considerably longer to claim a clean win for Claude: on the function of ARTs, on the priority dispute with Copenhagen, and on whether 21-hour discoveries like this one can be repeated rather than contested.