On September 23, Anthropic announced that its AI model Claude had discovered a new enzyme system, called ART, that contains repeat sequences reminiscent of the CRISPR gene-editing technology. Claude noticed DNA sequences and another gene next to a known enzyme that researchers had not previously described, and recognized them as parts of a single mechanism. Experiments also confirmed that multiple short RNAs are produced from the repeat sequences. It is still unclear whether ART will develop into a technology comparable to CRISPR. But the published technical report describes both the grounds for optimism and an important caveat: when the same search was run 10 more times, ART was not rediscovered.
A Mechanism Overlooked Next to a Known Enzyme
At the center of ART is a reverse transcriptase (RT), a type of enzyme that synthesizes DNA using RNA as a template. The preprint "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays," by Peter H. Yoon and colleagues at Anthropic, reports on an RT found in a jumbo phage, a virus with a large genome that infects bacteria, along with the structures surrounding it. The research has not yet been peer-reviewed.
The RT itself was already known from earlier studies. What Claude newly found was that a long array of repeated DNA sits next to it, along with a gene that appears to encode a partner protein. Anthropic named this combination array-associated reverse transcriptases, or ART for short. The AI did not design a new enzyme. It discovered a mechanism buried in existing data obtained from nature.
The path to the discovery is also interesting. Claude was initially tasked with finding combinations of an RT and an unknown partner gene. After examining the relationship between one candidate and its neighboring gene, it concluded that the gene was not a dedicated partner. It nonetheless judged that the RT was worth further investigation, kept analyzing, and directly examined the DNA sequence upstream of a related RT. There, it reportedly noticed that similar sequences appeared repeatedly.
The discovery is notable because Claude did not abandon the investigation after rejecting its initial hypothesis. While looking for protein-coding genes, it turned its attention to a regularity in a region that does not code for proteins. Approaches that narrow candidates based on features researchers set in advance tend to miss anomalies that fall outside those criteria. Here, the agent switched the question it was pursuing partway through, examined the number and features of the repeats, checked them against known systems and the literature, and then reported to human researchers.
How Is ART Similar to CRISPR?
In ART, distinct sequences sit between highly similar short sequences. CRISPR also has repeat sequences interspersed with different sequences called "spacers." This structural similarity is one reason the discovery is drawing attention.
In CRISPR-Cas9, RNA specifies the location of the target DNA, and Cas9 cuts that DNA. As explained by the Nobel Foundation, the key 2012 result was showing that researchers could cut DNA at a chosen site by changing the RNA sequence. Research that began with repeat sequences led to a technology that can specify targets freely.
There are also clues in ART that go beyond structural resemblance. The research team analyzed public RNA data from phage SA1, which infects staphylococci, and found that ART's repeat sequences are actively transcribed into RNA. Fifteen minutes after infection, RNA derived from the repeats made up as much as 8% of phage-derived RNA. This is not a share of all RNA in the host cell.
In addition, in an experiment expressing SA1's ART in E. coli, multiple short RNAs were confirmed to arise individually from the repeat sequences. In other words, the work has progressed from discovering a distinctive DNA sequence to experimentally confirming that it actually produces multiple RNAs.
The presence of repeat sequences and short RNAs has been confirmed in ART, but the function seen in CRISPR-Cas9, in which RNA specifies a target and cuts DNA, has not yet been demonstrated. When comparing the two, structural similarity must be kept separate from experimentally confirmed function.
| Point of comparison | CRISPR-Cas9 | ART (this study) |
|---|---|---|
| Sequence features | Repeats alternate with distinct spacers | Repeats alternate with distinct sequences, but length and conservation differ from CRISPR |
| What is known about the RNA | Functions as RNA that specifies the target DNA location | Multiple short RNAs arise from the repeat sequences |
| Relationship to the enzyme | Target specification by RNA and cutting by Cas9 are demonstrated | Not yet shown whether the RT is active or whether the short RNAs are its substrates |
| Stage as a technology | Method of cutting targeted DNA by changing the RNA is established | Still at the stage of clarifying the biological function |
The comparison is based on pages 6–8 and 13 of Anthropic's technical report and the Nobel Foundation's explanation of CRISPR-Cas9. It lines up the evidence available as of September 2026 for ART research and for the already established CRISPR-Cas9. It is not a comparison of performance measured under the same conditions.
If a single enzyme system can be supplied with multiple different RNAs, the enzyme's activity might change depending on the RNA. That is the hypothesis the research team is pursuing. However, the paper states explicitly that it has not yet been shown that ART's RT is enzymatically active, or that the RT interacts with its partner protein. No Cas gene has been found near ART either. Having a sequence structure similar to CRISPR does not justify assuming it also has the ability to cut DNA.
In scientifically evaluating the potential for a "CRISPR-class" technology, the next key milestone is not to look for more similarities with CRISPR. It is to show experimentally how ART's behavior changes when the RNA sequence is changed. Only once it is known that targets and reactions can be controlled can its potential as a technology be discussed concretely.
21.5 Hours of Searching, and the Experiments Done by Humans
The model used for the search was Claude Mythos 5. According to the technical report, it searched about 1.9 billion protein clusters and extracted about 200,000 RT clusters. Across 119 tasks, 949 agent sessions were run in total. Computational searching took 21.5 hours, and 19 reports were ultimately compiled.
The figure "949" does not refer to the number of researchers or robots working simultaneously. It is the total of various agent sessions, including those that carried out the search itself, those that checked plans and results, and those that organized shared records. And 21.5 hours is the time spent on computational searching; it does not mean everything through experiments and writing the paper was completed in 21.5 hours.
Nor does the number of candidates translate directly into the number of discoveries. Of the 17 protein families examined in detail as possible RT partners, new associations were confirmed for three. The remaining 14 were excluded or put on hold after being judged to be gene annotation errors, known systems, or genes that were merely nearby.
This narrowing shows the danger of measuring the value of research-support AI only by how many hypotheses it generates. To keep candidates worth taking to experiments, it is essential to doubt one's own explanations and discard candidates while checking them against existing knowledge. ART came out of that process.
Anthropic explains that at the molecular biology laboratory newly established in the Bay Area, all experimental work is done by human scientists. Claude searches sequences and reports candidates, and researchers review them before moving on to experiments. The autonomy of AI-driven searching should be considered separately from making the entire research process unmanned.
Ten More Runs Failed to Rediscover It: What This Reveals
The technical report states that when the same task and agent execution system were used for 10 additional searches, none of the runs discovered ART's repeat sequences. In nearly all of the runs that completed a broad search, the region containing ART was itself picked up as a candidate. In two of them, the agents followed that lineage in more detail. Even so, they did not examine the DNA sequence upstream of the RT, and so did not arrive at the discovery of ART.
The research team cites the very broad search space, and the fact that the agent does not behave in exactly the same way each time, as reasons. This does not mean that an experiment confirming the existence of short RNAs failed to be reproduced. The reproducibility of ART as a biological phenomenon and the reproducibility of the search, meaning whether an AI can arrive at the same discovery without hints, are separate issues.
Confusing the two could lead to a misjudgment. The failure to find ART in the repeat searches does not negate the ART sequences and RNAs that have already been confirmed. Conversely, a single successful discovery does not justify claiming that the same approach can reliably mass-produce new discoveries.
After the discovery, the research team also ran a more limited test in which ART sequences were given directly to models. Several top models reportedly recognized and explained the repeat structure at a high rate when they could read the target sequence directly. But a task that asks a model to recognize features after being told where the clue lies demands different capabilities from one that asks it to find that location within a huge database.
The ART case shows that there is a large gap, one that shapes scientific discovery, between the ability to recognize features when looking at data and the ability to choose which data to look at. Even if a model's ability to understand sequences improves, it leads to results only if the model examines that sequence during its search. Just as the decision to keep investigating after rejecting the initial hypothesis led to the discovery here, the design of the search matters as much as the model's knowledge: which questions to pursue, how far to investigate, and when to stop.
To determine whether ART could rival CRISPR, researchers first need to clarify how the enzyme system itself works and check whether the reaction can be controlled by changing the RNA. On the AI research side, the question is whether a model can repeatedly find unknown mechanisms across different searches without being told in advance where the answers are. Only when both a new technology that can be used in experiments and a search method that can keep finding such technologies are established can this single discovery grow into a result that changes how biological research is done.