Anthropic says its AI model Claude has autonomously discovered a previously uncharacterised enzyme system in the DNA of bacteriophages — viruses that infect bacteria — with properties that remind its scientists of CRISPR, the gene-editing technology that earned a Nobel Prize and launched a medical revolution. The company has named the system ART: array-associated reverse transcriptases.
The claim, announced by Anthropic on 23 September 2026 and covered by Al Jazeera, The Verge and Gizmodo, is the first published result from the company’s newly built wet lab in the San Francisco Bay Area. It is also, deliberately or not, a marketing event: Anthropic wants the world to believe Claude can do the noticing that scientific discovery is made of.
What Claude actually did
Anthropic gave Claude a broad prompt: search a massive database of DNA sequences for interesting examples of reverse transcriptases — enzymes that copy RNA into DNA. From there, the company says, its involvement was limited to the initial prompt and the lab work. Roughly 950 Claude agents combed the database for 21 hours, consuming about 210 million tokens, gathering more than 200,000 reverse transcriptases and generating some 3,500 candidate systems for review. One agent flagged something odd: a repeating pattern of DNA sequences sitting next to the gene for an unusual reverse transcriptase, in a jumbo phage.
Anthropic’s scientists tested the candidate in the lab and concluded the pattern marked a system that had never been characterised. The underlying reverse transcriptase, found in a jumbo phage, was already known to science. What is new, as QUT computer scientist Dimitri Perrin explained in The Conversation, is the recognition that it may be part of a larger system — a reverse transcriptase plus an accessory protein of unknown function plus an array of repeated non-coding DNA sequences, a combination that in a handful of other systems has turned out to be programmable DNA-cutting machinery. That is the architecture CRISPR has. What ART does is unknown.
The CRISPR comparison, and where it strains
The CRISPR framing is doing a lot of work in this announcement. CRISPR-the-natural-system is a bacterial immune mechanism; CRISPR-the-technology is a Nobel-winning gene-editing platform that has already delivered approved therapies. Multiple researchers told Al Jazeera the leap from one to the other should not be assumed. Kevin Blake, a microbiologist at Washington University School of Medicine, put it bluntly: “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.” Feng Zhang, one of CRISPR’s pioneers, called the identification of RNA-repeat arrays associated with reverse transcriptases “genuinely intriguing” and worth further investigation — encouragement for the method, not the headline.
Anthropic itself concedes the function, utility and significance of ART are unclear. Dario Amodei announced the discovery minutes before a UN Security Council meeting on AI safety, in a post suggesting ART “could represent a new gene editing mechanism” — the kind of timing that tells you what the announcement is for. The company is preparing for a public listing and is recruiting scientists into its life-sciences arm, and its rivals are not standing still: OpenAI has its own life-sciences push.
The part that actually matters
Strip away the CRISPR comparison and the durable news is the method. Anthropic demonstrated an end-to-end loop in which AI agents read the literature, mine enormous public datasets, propose candidates with written arguments for why each one matters, and humans test the survivors at the bench. The human role shifted from searching to judging. In Anthropic’s own description, Claude’s hypotheses are so prolific that the company now studies which of its own proposals the model judges worth testing — trying to teach the model its scientific taste.
That workflow is the plausible product, more than any single enzyme. A discovery engine that can survey 1.9 billion protein clusters — the scale Anthropic says the campaign covered — is a different research instrument from a search engine, and every large pharma and biotech lab will want one pointed at its own unsolved problems. The catch is verification: a candidate that survives AI review still has to survive a wet lab, and the wet-lab part remains entirely human. If AI systems are going to accelerate biology the way Anthropic’s Dario Amodei has predicted — cures within five to ten years — the bottleneck is not hypothesis generation any more. It is everything downstream of it.
New Zealand’s own life-science sector should read this as an early signal about tooling rather than a headline about a rival to gene editing. The labs that win the next decade will not be the ones with the biggest AI models; they will be the ones with the tightest loop between machine-scale search and human-scale verification. That loop is now a product category, and this is its first public proof point — with all the caveats that a company grading its own homework demands. As of publication, the underlying pre-print has not been through peer review, and no independent lab has replicated the characterisation of ART.
Anthropic, to its credit, published the system’s defining features before knowing what it does — a transparency choice that also lets the world check its work. What ART turns out to be will take years. What it says about how biology research will be organised — agent swarms reading the genome’s unlabelled dark matter while humans handle judgement and the bench — is already here.
