Everything About AI
NextFin News - Anthropic on Wednesday announced that its Claude artificial intelligence model had autonomously discovered a previously uncharacterized enzyme system in the DNA of bacteriophages, a finding the company said is reminiscent of CRISPR and could "represent a new gene editing mechanism." Within hours, scientists were drawing a sharp line: the system is CRISPR-like in architecture, but there is no evidence yet that it is CRISPR-like in function.
The announcement, released alongside a technical preprint that has not been peer-reviewed, is the first public result from a molecular biology group and laboratory Anthropic quietly assembled in the San Francisco Bay Area in the spring of 2026. It lands at a moment when the frontier AI labs are racing to prove that agentic systems can do more than answer questions — that they can notice something odd in nature and pursue it without being told what to look for. The scientific reaction suggests the bar for that claim is higher than the company's framing implies.
The Discovery: What Claude Actually Found
Anthropic named the system array-associated reverse transcriptases, or ART. It has three parts: a reverse transcriptase enzyme, which copies RNA into DNA; a neighboring partner gene of unknown function; and a long array of evenly spaced, non-coding DNA repeats. The enzyme itself was not new — it had been identified in previous studies, in a jumbo phage. What appears to be new is the recognition that the enzyme sits inside a larger system, complete with a repeat array and an accessory protein.
The layout recalls CRISPR, the bacterial immune system that became the foundation of modern gene editing. In CRISPR, an array holds a bank of RNA guides that point the system at specific DNA targets. ART arrays, according to the preprint, hold 3 to 21 copies of a short repeat. But ART lacks the cas genes that CRISPR systems carry nearby — the genes that encode the DNA-cutting machinery. And while Anthropic's human scientists confirmed in the lab that the ART array is read out as a set of distinct short RNAs, the preprint states plainly that the team has not yet shown the enzyme is active, or that it works on those RNAs.
In published data from a Staphylococcus phage, those short RNAs made up as much as 8% of the phage's total RNA just 15 minutes after infection — a signal strong enough to suggest the array is doing something biologically meaningful. Anthropic argues that only a handful of known systems share this combination of features, and that all of them are programmable and capable of cutting, copying, or pasting DNA. The company's chief executive, Dario Amodei, wrote on X that researchers suspect ART "could represent a new gene editing mechanism," while adding that its "precise function, biotechnological utility (if any), or level of significance is not yet clear."
How the Search Ran — and Why It Did Not Repeat
The mechanics of the search are, in several respects, the more consequential part of the story. Anthropic gave Claude a single high-level research brief: comb through a database of 1.9 billion protein clusters and find interesting new reverse transcriptase systems. Roughly 950 AI agents then ran for about 21 hours, consuming approximately 210 million tokens. They recovered around 200,000 enzyme clusters, scored 3,564 candidate partner families, and filed 19 reports for human review. The broader search identified roughly 3,500 candidate systems and narrowed them to about 20 for detailed analysis.
ART came from a side path. One agent, reading raw DNA beside an unusual enzyme, wrote that it could see a repeat pattern "by eye." It counted the repeats, compared them against known systems, searched the literature, and flagged the finding. Human scientists then validated it in the lab.
Here the story takes a turn that matters as much as the discovery itself. Anthropic ran the same campaign ten more times. Not a single rerun read the DNA upstream of the enzyme, and all ten missed the array. In fixed tests, the company's four most capable models described the array in at least 90% of attempts when they were handed the DNA sequence directly. When the models had to work through files and tools, that rate fell to as low as 32% — often because the model never read enough DNA to see a full repeat.
"At this point in time, we can say that ART is CRISPR-like in its architecture, but there is no evidence that it is CRISPR-like in its function," said Dimitri Perrin, who leads the School of Computer Science at the Queensland University of Technology and has worked on CRISPR and AI projects.
Perrin also noted that the enzyme itself was already known, and that the preprint offers no comparison showing whether a conventional bioinformatics pipeline would have found the same pattern. He raised a second question: of the roughly 3,500 candidate systems the agents identified, and the 20 selected for detailed analysis, it is unknown how many represent genuinely novel biology. Finding one intriguing result after exploring thousands of possibilities is not the same as demonstrating a system that reliably recognizes discoveries.
The Cautious Consensus
That distinction is where much of the external scientific reaction has settled. Kevin Blake, a microbiologist at Washington University School of Medicine, said CRISPR in nature — a bacterium's immune system — is very different from CRISPR-the-technology, the programmable gene-editing tool that won the 2020 Nobel Prize in Chemistry for Emmanuelle Charpentier and Jennifer Doudna.
"There's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any kind of therapeutic or practical application," Blake said. He added that countless CRISPR-like sequences remain uncatalogued because millions of bacterial species have yet to be studied.
The caution cuts against the headline framing. Anthropic's announcement leaned on the resemblance to CRISPR — the Nobel Prize-winning breakthrough that allowed scientists to edit the human genome — and Amodei's suggestion that ART could become a new gene-editing mechanism. The company is not alone in that optimism. Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute who reviewed the preprint, called the finding "an exciting example of how AI agents can contribute to biological discovery" and said the identification of RNA-repeat arrays associated with reverse transcriptases "is genuinely intriguing and merits further investigation." Stanley Qi, an associate professor of bioengineering at Stanford University, described the work as "incredibly exciting," praising the system's ability to recognize an unusual biological pattern and pursue it comprehensively.
But the gap between architectural resemblance and functional equivalence is wide, and it is the gap in which the commercial promise either materializes or evaporates. CRISPR's value as a technology rests on one property: the version found in nature could be tweaked to make programmed, targeted DNA edits. Whether ART can be programmed at all remains an open question. Anthropic says further experiments are underway to determine what the system does.
The Real Story: Autonomy, Not the Enzyme
Several researchers made the point that what was found may be less interesting than how it was found. The individual techniques involved are not fundamentally new: clustering sequences, detecting motifs and repeats, examining which genes occur together, and calculating how RNA folds are all established bioinformatics methods. A purpose-built conventional pipeline could probably have detected the same pattern.
What is different is autonomy. Claude was given a broad scientific objective and then chose which analyses to perform, which leads to follow, and which result was interesting enough to escalate for experimental validation. One agent followed an unexpected observation without being instructed to. That is a shift from AI as a tool that answers specific questions to AI as a system that decides where to point the tools.
"To me, that's the opportunity: scaling scientific attention," said Bo Wang, chief artificial intelligence scientist at the University of Toronto's health network.
"Nature contains an enormous diversity of molecular systems we barely understand, and some patterns are very complicated but highly meaningful; AI could greatly expand our ability to explore them more effectively and rapidly, in this case, in just 21 hours," said Stanley Qi.
This is the structural shift underneath the cyclical hype. The discovery itself may or may not become a practical tool — that is a cyclical question that will be answered by whether ART's function proves programmable. But the method represents a potential regime change in how biological discovery is organized: models that can survey datasets no single lab can exhaust, generate hypotheses at scale, and prioritize candidates for human testing.
Why the Reproducibility Finding Matters More Than the Headline
The ten failed reruns deserve more attention than they have received. Science rests on reproducibility: if a finding is real, another researcher following the same method should be able to reach it. Anthropic's result was not reproduced by its own system ten consecutive times. The company attributes this to the size of the search and the agents' unpredictable behavior — and it has built tools to study what Claude was "thinking" as it read the sequence, finding that two internal signals fired on the repeats just before the agent named them.
If agentic discovery is genuinely non-deterministic, that creates a problem no amount of enthusiasm solves. A discovery pipeline that cannot reliably reproduce its own hits cannot be audited, cannot be trusted at scale, and cannot form the basis of a drug-development program where a missed target has human consequences. This is not a cyclical obstacle that better prompting will fix; it is a structural question about whether stochastic agents can be made to satisfy the reproducibility norm that underpins experimental science. Either the field develops methods to make agentic search deterministic enough for validation, or agentic discovery remains a source of interesting leads rather than reliable findings.
The Competitive and Commercial Stakes
The announcement also fits into a widening race among frontier AI labs to move from purely computational work into wet-lab biology. Anthropic acquired Coefficient Bio for approximately $400 million in April 2026 and launched Claude Science, a bioinformatics platform, in June. OpenAI launched GPT-Rosalind, a model for life sciences research, in the same month. Anthropic's head of life sciences, Eric Kauderer-Abrams, has said the project will not run clinical trials and that the company is not competing with pharmaceutical and biotech firms that bring drugs to market — a deliberate boundary that keeps the work outside the strictest regulatory scrutiny while the company builds capability.
The company's big pharma clients — including Roche, Genentech, and Bristol Myers Squibb — may be reassured by that framing. But the ART project gives Anthropic access to capabilities beyond the reach of most academic labs: a team of six named human scientists, a private wet lab operating at biosafety levels 1 and 2, and models that can search 1.9 billion protein clusters in a day. Some academics have already raised concerns about whether tools like Claude Science will force researchers to choose between funding AI compute and funding trainees.
The moat is shifting. If agentic discovery proves reliable, competitive advantage in biology will accrue not to whoever has the best model alone, but to whoever combines the model with a wet lab, proprietary datasets, and the compute to run thousands of agent-hours on a single question. That is a capital-intensive moat, and it favors the frontier labs over individual researchers.
Cyclical or Structural: The Call
The right reading separates two forces that the announcement blends together. The hype around "the next CRISPR" is cyclical: it will mean-revert until ART's function is demonstrated, and the skepticism from Blake and Perrin is the market's equivalent of a correction toward fundamentals. History is littered with CRISPR-like systems that never became technologies. The evidence for a cyclical read is straightforward: the enzyme was already known, the function is unknown, the finding did not reproduce across ten reruns, and the preprint has not been peer-reviewed.
The autonomy shift, by contrast, is structural. A system that can survey more sequence space than any lab, notice an anomaly, and escalate it for validation changes the economics of discovery even if this particular anomaly proves biologically inert. The evidence: the search covered 1.9 billion protein clusters in 21 hours; the agent followed an unscripted lead; and the company has institutionalized the capability with a dedicated lab and life sciences organization. That does not revert on its own.
The strongest counter-thesis comes from Perrin: a conventional bioinformatics pipeline could have found ART, and the one-hit rate across thousands of candidates does not prove reliable discovery. If an independent lab reproduces a novel, functionally validated discovery with a standard pipeline at a comparable hit rate — or if Anthropic fails to produce a second validated discovery from the same approach within 12 to 18 months — the structural claim weakens materially. The falsifying signal is concrete: a peer-reviewed demonstration that ART's array is not programmable and that the enzyme performs no targeted DNA operation, combined with no additional agentic discoveries reaching experimental validation over the next year, would confirm that this was a curiosity rather than a regime change.
What to Watch
In the short term, watch for the peer-review outcome of the preprint and any functional characterization of ART — specifically whether the enzyme can be shown to act on the short RNAs the array produces. In the medium term, watch whether Anthropic or a competitor reports a second agentic discovery that reaches experimental validation, which would move the method from anecdote to pipeline. In the long term, watch the capital allocation: whether frontier labs continue to build private wet labs and whether academic researchers gain comparable access or face a widening capability gap.
The base case is that ART's function is eventually characterized, that it resembles other programmable systems in architecture but may not match CRISPR's programmability, and that the agentic method produces a steady trickle of leads rather than a flood of validated tools. The upside case is that ART becomes a genuinely programmable system, joining the small set of natural molecular machines that became biotechnology. The downside case is that ART proves to be one more CRISPR-like curiosity, and the reproducibility problem keeps agentic discovery in the lead-generation stage for years.
The discovery is real; the comparison to CRISPR is, for now, architecture without evidence of function. The market for AI-driven biology is pricing a regime change in how science gets done — and that part may be right even if this particular enzyme never cuts anything.
Explore more exclusive insights at nextfin.ai.
