Anthropic said Claude identified a novel enzyme system with features reminiscent of CRISPR after screening DNA data from bacteria-infecting viruses. This is an early biological discovery, not a proven gene-editing tool. What is noteworthy is how the research team combined AI with laboratory experiments to filter out a testable hypothesis.
According to Anthropic’s announcement, the system was named array-associated reverse transcriptases (ART). It consists of a reverse transcriptase enzyme, an adjacent gene, and an array of regularly spaced DNA repeats. This repeat array reminded the research team of CRISPR. Initial experiments showed that the ART repeats are expressed into distinct short RNA molecules, but their function remains undetermined.
Anthropic described Claude screening more than 200,000 reverse transcriptase enzymes, selecting 3,500 candidate systems, and then narrowing them down to 20 cases for in-depth analysis. During a 21-hour search run, around 950 agents processed a total of 210 million tokens. One agent recognized a DNA repeat pattern next to an unusual enzyme, then checked the spacing between repeats, compared it with known systems, and searched the literature to see if this pattern had ever been described.
These figures were self-reported by Anthropic. They demonstrate the scale of the search organized by the team, but do not by themselves prove the scientific quality of the discovery. The preprint published by the company needs further review and peer evaluation by the research community. Anthropic also stated that the underlying reverse transcriptase enzyme had appeared in previous studies; the contribution the company attributes to Claude is identifying the accompanying features, including the DNA repeat array and an accessory protein of unknown function.
Claude proposes, scientists test in the laboratory
Anthropic does not describe the process as handing off the entire study to a model and receiving complete conclusions. The team started with a search prompt, allowing Claude to analyze data and propose candidates. Humans reviewed the reports, eliminated most unconvincing hypotheses, and only then took promising candidates into experiments. In the case of ART, scientists carried out wet-lab work and continued analyzing the results alongside Claude.
This division of labor is important because biological data contains many patterns that look noteworthy but do not necessarily lead to an actual function. The fact that a system is found in data does not mean we know what it does. Anthropic says researchers are continuing to investigate ART’s primary function; therefore, this discovery should not yet be interpreted as a new gene-editing method or a ready-to-use medical application.
The company said its Bay Area lab works only at biosafety levels BSL-1 and BSL-2, does not handle pathogens capable of infecting humans, and all experimental procedures are performed by humans. Meanwhile, CEO Dario Amodei has previously voiced concerns about AI potentially assisting dangerous biological activities. A TechCrunch article noted this very contrast: Anthropic is pursuing the potential benefits of AI in research while keeping humans in the loop for current experiments.
This is Anthropic’s own statement about its work. An outside scientist, Feng Zhang, described finding repeat RNA arrays associated with reverse transcriptase as interesting and worthy of further study, without confirming any applications for ART. That distinction should be maintained when evaluating the news: AI can help humans find candidates faster, but the biological value of those candidates still requires independent verification.
If this approach holds up in subsequent studies, the replicable part may not necessarily be a single discovery, but rather the workflow of using agents to screen hundreds of thousands of samples, generate hypotheses, and allocate lab resources to a select few candidates. But to know whether that process genuinely accelerates research while maintaining reliability, we will have to wait for further verified results and implementations outside of Anthropic.