For years, tech executives have promised that artificial intelligence would revolutionize biology, compressing decades of discovery into days. Last week, that promise seemed to take a tangible step forward when Anthropic announced that its large language model, Claude, had identified an enzyme system with properties “reminiscent of Crispr,” the Nobel Prize-winning gene-editing tool. According to the company, roughly 950 Claude agents working in parallel combed through vast genomic databases and, in just 21.5 hours, landed on a set of genetic sequences that looked strikingly like the repetitive arrays found in CRISPR systems. The announcement was met with a mixture of genuine admiration and careful skepticism. Fyodor Urnov, a gene-editing expert at the University of California, Berkeley and director for therapeutic R&D at the Innovative Genomics Institute, offered a sincere compliment: “I sincerely compliment Anthropic for telling the world about their discovery.” But other scientists, while intrigued, were quick to point out that this is not yet a finished story. The discovery is raw, unvalidated, and far from being a proven tool. It is a tantalizing clue, not a breakthrough cure. And as with all scientific findings, the real test begins only after the press release fades and the hard work of lab verification begins.
Anthropic, best known as an AI safety company, has been quietly building a wet lab for drug discovery, and this discovery is the first to emerge from that effort. In a technical report that has not yet been peer-reviewed, the company described how Claude was asked to search through huge genomic databases for “interesting new examples” of reverse transcriptases—proteins that copy RNA into DNA, a process that reverses the usual cellular flow of genetic information. While humans typically go from DNA to RNA to protein, many organisms, especially viruses and bacteria, use reverse transcriptases for a variety of biological functions, including inserting segments of DNA into genomes. The AI agents initially identified more than 200,000 possible reverse transcriptases, then filtered those down to several thousand that appeared to be novel, and finally zeroed in on an “unusual” reverse transcriptase family that contains a long region of repeated DNA sequences. Anthropic named this system ART, short for array-associated reverse transcriptases, and found it in jumbo phages—large viruses that infect bacteria. The company was candid about its limits, posting on X that “we don’t yet understand what this system does,” while noting that only a handful of known systems share its features and that all of them are able to cut, copy, and paste DNA. That honesty is welcome, but it also underscores how much remains unknown. The system may be fascinating, but it could also be a dead end for gene editing. The leap from a promising genetic sequence to a reliable, safe, and useful therapeutic tool is enormous, and no algorithm can make that leap alone.
The moment of discovery, as described by Anthropic, is almost cinematic in its artificial intelligence. One of the Claude agents, scanning through the data, apparently paused in astonishment and wrote: “I can see by eye a tandem repeat array … that’s a Crispr-like … repeat array?!” The agent also acknowledged, in a moment of scientific honesty, that this system could be a retron—a type of bacterial immune system that is related to Crispr but not identical to it. Both Crispr and retrons are ancient immune defenses that bacteria use to fend off viruses, and both have been harnessed by scientists for gene editing. But they are not interchangeable. Crispr has become a versatile multi-tool because it can be programmed to cut specific DNA sequences with remarkable precision, making it invaluable for genetic research and, increasingly, for medicine. Retrons, on the other hand, are less well known and less powerful, though they have been used in some gene-editing applications. So when Anthropic’s blog post leaned heavily on the Crispr comparison, some researchers felt the company was getting ahead of itself. Le Cong, a professor at Stanford University who studies the integration of AI into genome engineering, put it bluntly: “The experiments are still in the queue. The PR is already live.” Cong offered a vivid analogy for the situation: imagine standing on Santa Monica Beach and scanning through all the sand to find a diamond. The AI spots something shiny, but you have to go back to the lab to determine whether it is glass or a diamond. He was not dismissing the finding, but he was reminding everyone that a shiny object is not yet a gem.
Seth Shipman, an associate investigator at the Gladstone Institutes, went a step further in his assessment. He does not believe that what Anthropic found is actually a new Crispr system. But he still finds the discovery interesting—not because of what it is, but because of how it was found. “The novel thing is how they found it, not what it is,” he said. Shipman’s lab has used retrons to build gene-editing systems, so he knows firsthand that retrons can be adapted for such purposes. That possibility is real, but it is also limited. The discovery is an early-stage computational prediction, and one physical experiment was included in the technical report. That is a start, but it is not enough to say that ART is a usable gene-editing tool, let alone a new Crispr. To go from a sequence to a system, scientists need to characterize the enzyme, determine its natural function, and then figure out how to repurpose it for human applications. They need to test it in living cells, measure its efficiency and accuracy, assess its safety, and compare it to existing tools. This is painstaking, expensive, and slow work, but it is essential. The AI can point the way, but it cannot do the benchwork. As Urnov noted, Anthropic deserves credit for sharing its discovery with the world, but even he would agree that sharing a discovery is not the same as proving one. The scientific community’s response is a healthy reminder that progress in biology is measured not in press releases, but in reproducible experiments and peer-reviewed results.
Beyond the specific claims and counterclaims, this episode captures a broader moment in the relationship between artificial intelligence and science. AI models are no longer just tools for parsing text or generating images; they are being used to generate hypotheses, design experiments, and explore vast spaces of biological possibility that human researchers could never manually search. The idea of having hundreds of AI agents working around the clock, each independently exploring genomic data and then coming together to share findings, is genuinely exciting. It suggests a future where AI accelerates the pace of discovery not by replacing scientists, but by acting as tireless research assistants that can spot patterns and peculiarities that human eyes might miss. At the same time, this future comes with risks. When an AI discovery is announced with bold language and a splashy blog post before the science has been validated, it can create confusion and overhype. It can lead the public to believe that a miracle cure is just around the corner, or that a tool like Crispr has already been surpassed. It can also put pressure on other researchers to respond, validate, or debunk claims on an accelerated timeline, before the necessary experiments have been done. The best AI-driven science will be the kind that combines computational power with old-fashioned scientific rigor: AI proposes, human researchers dispose. In that sense, Anthropic’s announcement is a useful case study, not just for biologists, but for anyone who cares about how technology is changing the way knowledge is created.
So where does this leave us? We have a newly discovered family of reverse transcriptases, named ART, found in jumbo phages by an AI that was guided by human curiosity. It has a repetitive structure that looks, at least superficially, like the arrays that define Crispr. It might be a retron, a Crispr-like system, or something entirely different. It might one day become a useful gene-editing tool, or it might remain an interesting footnote in the genomic history books. No one knows yet. What is clear is that Anthropic is not claiming to have reinvented Crispr; it is claiming to have found something that reminds it of Crispr, and that deserves further investigation. That is a reasonable and even modest claim, and it is exactly the kind of claim that science is designed to test. The AI did its part, scanning millions of sequences and surfacing a candidate that human researchers would not have found so quickly. Now the humans must do theirs, in the wet lab, with pipettes and petri dishes and months of patient experimentation. If ART proves to be just glass, the exercise will still have value, because it shows that AI can be a powerful compass for navigating the unknown. If it turns out to be a diamond, the world will be better for it. Either way, the story of this discovery is not really about a machine outperforming humans. It is about a partnership—messy, imperfect, and full of promise. The AI found the shiny thing. The humans will tell us what it actually is.
