# Anthropic Says Claude Has Helped Make a Biology Discovery, and It Looks a Bit Like CRISPR > Source: > Published: 2026-09-23 19:17:55+00:00 Anthropic’s [biology lab](https://officechai.com/ai/amid-voicing-non-stop-concerns-about-ai-safety-anthropic-sets-up-physical-biology-lab/), which was largely unknown to the general public until last week, is already producing results. Anthropic has said that Claude, its AI model, has discovered a previously unknown biological system with features reminiscent of CRISPR, the gene-editing technology that earned its inventors a Nobel Prize. The announcement came alongside the launch of a new life sciences research group and a wet laboratory in the Bay Area. The discovery is early-stage, and Anthropic is upfront that it doesn’t yet know what the system actually does. But the way it was found is the more interesting story for anyone following AI. ### First, a quick biology primer Some of the most important tools in modern biology began as odd observations in nature. Restriction enzymes, which cut DNA at specific sequences, were found in bacterial defenses against viruses. Researchers realized they could use them to cut and splice genes, which helped launch the biotech industry. Taq polymerase, an enzyme from a bacterium living in a Yellowstone hot spring, became the basis of PCR, the DNA-copying technique behind much of modern diagnostics. CRISPR started as a strange repeating pattern in bacterial DNA and is now the foundation of gene-editing medicines. The pattern is consistent: someone notices something strange in the enormous variety of molecular machinery in nature, and it turns into a tool. Anthropic’s new research group is betting that AI can speed up that process. ### What Claude found The team pointed Claude at a massive database of DNA sequences and asked it to hunt for interesting new examples of reverse transcriptases, or RTs. These are enzymes that copy RNA into DNA, and many bacteria use them as part of their immune systems. Anthropic says its scientists only wrote the initial prompt and did the later lab work. Claude did the searching. Roughly 950 AI agents worked for 21 hours, using 210 million tokens. They gathered more than 200,000 RTs, flagged 3,500 as new candidate systems, and narrowed those to the 20 most compelling, writing a human-readable report on each. Anthropic notes that this kind of analysis can take an expert scientist weeks to months. Then one agent noticed something odd next to the gene for an unusual RT: a repeating pattern in the DNA. According to Anthropic, the agent reacted in its notes with surprise, saying it could see a tandem repeat array by eye and asking whether it was a CRISPR-like repeat array. From there, the agent behaved much as a human scientist might. It counted the repeats, measured the spacing between them, compared the layout with known RT systems, and searched the literature to see whether anyone had reported the pattern before. Convinced it had found something new, it filed a report for human review. ### Why the “CRISPR-like” part matters In CRISPR systems, an array of repeated DNA sequences stores a library of short RNA guides. Those guides tell the molecular machinery where to cut, which is what makes CRISPR programmable and so useful as a tool. The new system, which Anthropic calls array-associated reverse transcriptases, or ART, has three parts: the RT, a partner gene beside it, and a long array of evenly spaced repeats. It appears mainly in bacteriophages, the viruses that infect bacteria. Early experiments show the array is expressed as a set of distinct short RNAs, which hints that something analogous to CRISPR’s guide system may be happening. Anthropic says this combination of features has only been seen together in a handful of other systems, all of them programmable and capable of operations like cutting, copying and pasting DNA. Several of those are now being developed as promising tools. One caveat: the underlying RT, found in a jumbo phage, had been identified in earlier studies. What Claude appears to be the first to notice is the surrounding array and the accessory protein of unknown function. ### What this doesn’t mean It’s worth being careful here. Anthropic has not shown that ART is a new gene-editing tool, and it says its work to understand the system’s primary function is ongoing. The company chose to share early, partly to demonstrate what Claude can do and partly to give the wider research community a look at its work. A pre-print has been released, meaning it hasn’t yet gone through formal peer review. Still, the discovery has drawn some notable attention. Feng Zhang, a pioneer of CRISPR genome editing at MIT and the Broad Institute, reviewed the pre-print and called it an exciting example of AI agents contributing to biological discovery. He described the identification of RNA-repeat arrays associated with reverse transcriptases as “genuinely intriguing” and worthy of further investigation. ### How the lab works Anthropic’s approach pairs AI-driven analysis with conventional bench science. Claude surveys a protein family, reproduces known results from public data to check its own methods, looks for family members that fit no described system, and writes a short report proposing a function for each candidate. It then critically evaluates that evidence, and most candidates are eliminated at this stage. A survey may end with one candidate worth testing, or none. Survivors go to human scientists, who express the protein in standard lab strains and characterize it biochemically and structurally, with Claude helping interpret the data. The lab works only at the lower biosafety levels (BSL-1 and BSL-2) and does not handle pathogens that can infect humans. All lab work is done by people, not robots. There’s also a feedback loop. Because Claude generates hundreds to thousands of hypotheses in a single campaign, the team studies which ones its scientists judge worth testing and feeds that back into Claude’s instructions, in effect teaching it their scientific taste. ### Why it matters The takeaway isn’t really about one enzyme. It’s about the shape of the workflow. Genome mining, the search for uncharacterized genes hiding in huge sequence databases, is exactly the kind of task where the bottleneck has been human attention: someone has to look through the data, notice something strange, and work out whether it’s real. Anthropic’s claim is that AI agents can now do that first, labor-intensive stretch at scale, with humans stepping in for judgment calls and experiments. Whether ART turns out to be a genuine breakthrough or a curiosity will depend on experiments still underway. But as a proof of concept that an AI model can autonomously spot an anomaly, chase it down, and hand scientists a lead worth testing, it’s a notable data point in the argument over how much AI can accelerate science.