Tech leaders have long promoted artificial-intelligence models as accelerators for biological research. On September 23 Anthropic released a statement that its large language model, Claude, had flagged an enzyme system whose properties it described as “reminiscent of CRISPR,” the Nobel-winning gene-editing technology. The claim marks the first public output from Anthropic’s biology-focused research group formed earlier this year.
Scientists expressed caution about the announcement. Stanford professor Le Cong, who studies AI-enhanced genome engineering, noted that experimental validation is still pending while the press release is already public. Anthropic’s own X post admitted that the function of the newly identified system is unknown, and that only a handful of known systems share its cut-copy-paste DNA capabilities. The technical report describing a single wet-lab experiment has not undergone peer review.
Claude was tasked with scanning massive genomic repositories for “interesting new examples” of reverse transcriptases, enzymes that convert RNA into DNA. The model initially flagged more than 200,000 candidates, then filtered down to several thousand that appeared novel, and finally isolated an unusual family containing a long repeat region reminiscent of CRISPR arrays. Anthropic has named the discovery ART, short for array-associated reverse transcriptases, and reports that it occurs in jumbo phages, large viruses that infect bacteria.
The AI agent described the repeat region as a “Crispr-like … repeat array” and also considered the possibility that the system is a retron, another bacterial immune mechanism. Retrons have been repurposed for limited gene-editing applications but lack the multifunctional capabilities of CRISPR. Anthropic’s blog post highlighted the CRISPR analogy, a framing that some observers view as overstated given the current uncertainty about the system’s function.
Seth Shipman of the Gladstone Institutes emphasized that the novelty lies in the discovery method rather than the enzyme itself. His laboratory has employed retrons to construct gene-editing tools, suggesting that ART could be harnessed similarly. Shipman praised the speed of Claude’s search,compressing months of manual database mining into a single day,but warned that scientists, not the model alone, drove the investigation.
The reverse transcriptase at the center of ART was already reported by microbiologist Jason Gill and colleagues in a 2021 study of jumbo phages; Claude’s contribution was the identification of surrounding repeat elements that hint at a CRISPR-like architecture. Gill noted that AI excels at pattern detection but still requires a guiding hypothesis. Anthropic must now demonstrate any gene-editing activity, a step reminiscent of the decades-long path from CRISPR’s 1987 discovery to its first therapeutic approval in late 2023, underscoring the long timeline for translating basic findings into medicines.