Anthropic says its Claude agents have identified a possible Crispr-like biological system, but independent researchers caution that the discovery remains an early hypothesis rather than a demonstrated gene-editing breakthrough.
The company says about 950 agents working simultaneously found the relevant genetic sequences in 21.5 hours. The system, which Anthropic calls ART—short for array-associated reverse transcriptases—was identified in jumbo phages, large viruses that infect bacteria.
What Claude identified
Researchers asked Claude to search large genomic databases for unusual examples of reverse transcriptases. These are proteins that copy RNA into DNA, reversing the more familiar flow of genetic information in cells. Organisms use them for several biological functions, including inserting DNA segments.
The agents first located more than 200,000 possible reverse transcriptases. After filtering several thousand candidates that appeared to be new, they focused on an unusual family with a long stretch of repeated DNA. Those repeats resembled a Crispr array, prompting the model to flag the sequence as potentially significant.
Anthropic has stressed that it does not yet know what the system does. The company has also said that only a small number of known systems share its features, and that those systems can cut, copy and paste DNA. one physical experiment appears in Anthropic's technical report, but the report has not been peer-reviewed.
That distinction matters because the reverse transcriptase itself was identified previously. Jason Gill, a microbiologist at Texas A&M University, and colleagues described the enzyme in a 2021 paper about jumbo phages. The potentially new observation is that Claude identified surrounding repeats suggesting the enzyme could be part of a Crispr-like arrangement.
Scientists separate the method from the claim
Several researchers praised the speed of the search while questioning the comparison with Crispr. Seth Shipman of the Gladstone Institutes said the notable development may be the way the sequence was found, rather than the biological identity of the system. He added that retrons—another type of bacterial immune system—can be used in gene-editing research, although they are not the same as Crispr.
Le Cong, a Stanford University professor who studies AI and genome engineering, compared the result to spotting something shiny on a beach and then returning to the laboratory to determine whether it is valuable. His point is that computational pattern recognition can produce a candidate, while experiments must establish what the candidate actually does.
Gill said ART appears to have no obvious relationship to known Crispr systems and that Anthropic must demonstrate genuine gene-editing activity. He also cautioned against describing the result as an autonomous discovery by Claude, because scientists selected the research question, shaped the search and interpreted the output.
Shipman said finding new reverse transcriptases can take months of manual database work. Claude's ability to perform a related search in a day could therefore be useful for large-data biological projects. That does not establish that the model will accelerate every area of life-science research.
Questions about transparency and proof
The scientific community cannot fully assess what information was used to train the model. That has raised concerns about reproducibility, particularly because major journals increasingly expect researchers to publish code and methods openly. A researcher studying the same enzymes also questioned whether unpublished findings shared with the public version of Claude influenced the result. Anthropic said Claude was not trained on user transcripts and that its molecular biology team did not have access to them.
The path from biological discovery to medicine is long. Crispr sequences were first identified in bacteria in 1987, while Jennifer Doudna and Emmanuelle Charpentier demonstrated its programmable DNA-cutting potential in 2012. Crispr is now being tested in dozens of clinical trials for cardiovascular conditions, cancers, autoimmune diseases and rare disorders, but only one drug using the technology has reached the market, after approval in late 2023.
Anthropic CEO Dario Amodei has suggested that Claude could eventually control laboratory equipment and perform experiments autonomously. That possibility remains distant. Anthropic said its laboratories operate at the lowest biosafety levels and do not contain anything greatly harmful to humans.
For now, the next step is conventional laboratory validation. Scientists must determine whether ART is a functional biological system, whether it can edit genes and whether it offers advantages over existing tools. The discovery also raises a broader test for AI research: whether a model can find a genuinely useful biological mechanism without being guided toward a known type of answer.
Conclusion
Anthropic's result shows how AI can rapidly search enormous genomic datasets and identify patterns that merit investigation. It does not yet show that ART is a new Crispr system or a practical gene-editing tool. Laboratory experiments and independent scientific scrutiny will determine the finding's significance.
Frequently Asked Questions
Q. What did Anthropic's Claude agents find?
They identified ART, an unusual reverse transcriptase family with repeated DNA sequences resembling a Crispr array in jumbo phages.
Q. Is ART confirmed to be a gene-editing tool?
No. Its gene-editing activity and usefulness have not yet been established through sufficient laboratory testing.
Q. Was the reverse transcriptase completely new?
No. Jason Gill and colleagues had identified the same reverse transcriptase in a 2021 paper. The potentially new observation involved the repeats around it.
Q. How quickly did the AI search find the sequences?
Anthropic says about 950 Claude agents found the relevant sequences in 21.5 hours.
Q. Has Anthropic's technical report been peer-reviewed?
No. The report includes one physical experiment but has not been peer-reviewed.
Q. Why are scientists cautious about the Crispr comparison?
Researchers say the sequence may be a retron or another system, and experiments are needed to show whether it performs gene editing.
Q. What must happen next?
Scientists need to test ART in the laboratory and determine what the system does and whether it can be useful for gene editing.
Q. How does this compare with Crispr's development?
Crispr sequences were identified in 1987, and their programmable DNA-cutting potential was demonstrated in 2012.
