Hype, Skepticism, and the Fine Print Behind Anthropic’s AI-Hunted "CRISPR-Like" Gene Editing Breakthrough
For years, leaders across the tech industry have repeatedly framed AI models as a transformative force that will drastically speed up the pace of breakthrough biological discovery. Last month, they claimed to have delivered on that promise.
In a September 23 public announcement, AI developer Anthropic stated that its large language model Claude had identified a novel enzyme system with traits they described as “reminiscent of CRISPR,” the Nobel Prize-winning programmable gene-editing tool that upended life sciences research over the past decade. Per Anthropic, roughly 950 parallel-running Claude agents sifted through massive genomic databases to pinpoint the promising, previously little-documented genetic sequences in just 21.5 hours.
“I sincerely compliment Anthropic for telling the world about their discovery openly,” notes Fyodor Urnov, a leading gene-editing expert at the University of California, Berkeley, and director of therapeutic R&D at UC Berkeley’s Innovative Genomics Institute. The IGI has an ongoing research collaboration with Anthropic, but was not involved in this new discovery work.
Other domain experts are far less enthusiastic about the finding, which marks the first public output from a dedicated biology research team Anthropic launched earlier this year. “All the validation experiments are still in the queue, and the PR campaign is already live,” says Le Cong, a Stanford University professor whose work focuses on integrating AI tools into genome engineering research.
Anthropic, which has built a full in-house wet lab for drug discovery R&D, has been transparent that this preliminary finding is far from the end of their research on the system. “We don’t yet understand what this system does in full,” the company posted on X, “but only a handful of known biological systems share its core features, and every one of those systems is capable of cutting, copying, and pasting DNA segments.” A single small preliminary physical experiment is included with their associated non-peer-reviewed technical report, but far more lab work is required to establish the actual significance of the discovery. It is still entirely unconfirmed whether this enzyme system can function as a viable gene-editing tool at all, and if it can, whether it will offer any meaningful advantages over existing widely used gene-editing platforms.
“Let’s say we are on Santa Monica Beach and trying to scan through every grain of sand to find a diamond,” Cong explains. “AI found this thing that looks very shiny, and then you have to go back to the lab to figure out — is it just a piece of broken glass? Or is it actually the diamond we are looking for?”
Claude identified the shiny, promising object after Anthropic’s research team prompted it to comb through massive public genomic databases for “interesting new examples” of reverse transcriptases: a class of proteins that copy RNA sequences back into DNA, the reverse of the standard transcription process in human cells where DNA is converted to RNA. This reverse transcription process is used across many microbes for a range of biological functions, including inserting new DNA segments into host genomes.
Claude’s automated agents first flagged over 200,000 candidate reverse transcriptase sequences, per Anthropic, filtered that set down to several thousand sequences that appeared to be previously undocumented, then narrowed the scope even further to finally identify an “unusual” new reverse transcriptase family with a long stretch of repeated DNA sequences that closely match the structure of CRISPR arrays. Anthropic named this new system ART, short for array-associated reverse transcriptases, which they found in jumbo phages — extra large viruses that infect bacteria. One of the AI agents flagged the unusual pattern in its running notes, writing, “I can see by eye a tandem repeat array … that's a Crispr-like … repeat array?!” The agent also noted the system could potentially be a retron, another type of bacterial immune system.
CRISPR and retrons are both natural bacterial defense systems, but while retrons have some limited niche applications in gene editing, they are not the flexible, highly adaptable multi-tool that CRISPR turned out to be for research and clinical use. That makes Anthropic’s choice to heavily emphasize the CRISPR comparison in their public announcement far from surprising.
Seth Shipman, associate investigator at the Gladstone Institutes, does not believe the system Anthropic found is a new type of CRISPR, but says the discovery is still scientifically interesting. “The novel thing is how they found it, not what it is,” he points out. Shipman’s own lab regularly uses retrons to build custom gene-editing systems, so he confirms that this class of proteins can be engineered for editing use cases. Manually mining genomic databases to identify new reverse transcriptase variants typically takes human researchers months of work, so Shipman says Anthropic pulling off that screening process in less than a day is genuinely impressive, but he pushes back on framing the discovery as fully autonomous AI-generated work. “I think we have to be careful about saying that Claude autonomously discovered something, because there are experienced, skilled human scientists involved at every step of the study’s design and execution,” he notes.
In fact, this exact reverse transcriptase sequence had already been identified by Texas A&M University microbiologist Jason Gill and his co-authors back in a 2021 published paper focused on jumbo phage genomics. What Claude delivered that was new, was identifying the unusual repeated sequences surrounding the enzyme that hinted it might be part of a CRISPR-like array structure. “These models are good at finding patterns, better than a person staring at sequence data directly with their own eyes can be,” Gill says. “A person could do all this work, but you have to already have a specific hypothesis in mind to know where to look.” He adds that the ART system appears to have no obvious evolutionary relationship to any known CRISPR system, and the full burden of proof falls on Anthropic to demonstrate that the system actually has functional gene-editing activity.
Cong gives the Anthropic research team credit for identifying a perfectly suited problem that plays directly to Claude’s pattern-spotting strengths. The model may well prove useful for similar large-scale genomic screening projects, but that single success does not mean Claude will suddenly accelerate every category of life sciences discovery overnight. A major unresolved transparency issue also hangs over the work: the broader scientific community has no way to evaluate what full datasets Claude was trained on, making this finding unusually opaque at a time when top peer-reviewed journals require researchers to freely publish all underlying code and raw data to ensure research reproducibility. This lack of transparency even led one independent researcher who studies this exact class of enzymes to raise suspicions that the model was trained on his own yet-unpublished findings, which he had previously shared with the public consumer version of Claude. Anthropic pushed back on that claim in a statement to The New York Times, noting “Claude was not trained on any user transcripts, and our molecular biology team has no such access, either.”
Even if Claude had somehow identified the next major breakthrough in gene editing, human lab scientists would still need to test and validate every part of that hypothesis in wet lab work to prove it is functional and usable. In an X post discussing the discovery, Anthropic CEO Dario Amodei speculated, “Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment.” Even if that scenario does prove technically feasible, it will be a very distant future development, and immediately raises a host of unaddressed questions around biosecurity, safety oversight, and accountability. Amodei quickly followed up the comment to note that Anthropic’s in-house labs currently operate at the lowest biosafety containment level, and do not house any materials that pose high levels of harm to humans.
Major breakthroughs in this field do not arrive quickly. It took 25 years from the first initial spotting of unusual CRISPR repeated sequences in bacterial DNA in 1987, to the 2012 landmark study where Jennifer Doudna and Emmanuelle Charpentier demonstrated that the system could be programmed to precisely cut targeted DNA sequences. Today CRISPR is ubiquitous in research labs around the world, and is being tested in dozens of clinical trials for cardiovascular conditions, cancers, autoimmune diseases, and rare genetic disorders — but only one single CRISPR-based drug has earned full regulatory approval to date, as of late 2023. That long timeline is a testament to the slow, winding arc of scientific progress, and the massive difficulty of translating promising preliminary basic science findings into usable clinical treatments.
The limitations of Anthropic’s ART discovery also point to a far more fundamental, unresolved question for the entire AI for science field. What made CRISPR truly revolutionary was that it represented a completely unrecognized biological system no human scientist had even hypothesized could exist. It remains to be seen whether any AI model is actually capable of identifying a genuinely high-impact useful discovery without being given very specific, predefined targets to search for. “If you have an AI that only trained on all knowledge from before people ever discovered CRISPR, and then that AI actually discovered CRISPR entirely on its own, that seems to be a far better setup to test what a truly independent AI scientist could do,” Cong says.
