Level 1 — Absolute Beginner
Scientists at a company called Anthropic use an AI named Claude. On 23 September 2026 they shared a surprise. The AI found something new in DNA.
About 950 copies of the AI worked at the same time. They worked for 21 hours. They looked at more than 200,000 enzymes. Enzymes are small tools inside living things.
The AI found a new system. The scientists call it ART. It lives inside viruses that attack bacteria.
The system has an enzyme and a long row of repeating DNA. This looks like CRISPR, a famous tool. But nobody knows yet what ART does.
- scientist
- a person who studies the natural world
- DNA
- the material inside living things that carries instructions
- enzyme
- a substance in living things that makes chemical changes happen
- system
- a set of parts that work together
- virus
- a very small thing that can make people or cells sick
- bacteria
- very small living things, too small to see
- repeat
- something that happens again and again
- famous
- known by many people
Level 2 — Elementary
Anthropic, the company behind the Claude models, announced on 23 September 2026 that its AI had found a biological system that scientists had not described before. The company gave Claude only high level direction and let it decide how to search.
Around 950 Claude agents ran at the same time for 21 hours, using 210 million tokens of computing work. They examined more than 200,000 enzymes called reverse transcriptases, picked out about 3,500 possible systems, and reduced that list to the 20 most interesting ones.
The result is a system the researchers named ART, short for array associated reverse transcriptases. It appears mostly in bacteriophages, the viruses that infect bacteria. Each ART system has three parts: the enzyme itself, a partner protein, and a long array of DNA repeats spaced evenly apart.
That structure is what excites biologists, because CRISPR has a similar pattern of repeats, and early experiments suggest the ART array is turned into a set of short RNA pieces. Anthropic has published a preprint and invited other scientists to help work out what the system actually does.
- announce
- to tell the public about something officially
- biological
- relating to living things
- examine
- to look at something carefully
- array
- a row or arrangement of things in order
- protein
- a large molecule that does work inside living cells
- structure
- the way the parts of something are arranged
- experiment
- a test done to learn something new
- preprint
- a scientific paper shared before formal review
Level 3 — Intermediate
Anthropic described on 23 September 2026 an unusual result: a biological system that appears to be new to science, surfaced not by a laboratory team following a hypothesis but by roughly 950 copies of its Claude model working in parallel with only high level direction about where to look.
The search itself was industrial in scale. Over 21 hours the agents consumed 210 million tokens while combing genetic databases for reverse transcriptases, enzymes that copy RNA back into DNA. From more than 200,000 candidates they flagged about 3,500 systems worth attention and then narrowed that to the 20 that looked most unusual.
The system they settled on has been named ART, for array associated reverse transcriptases, and it turns up chiefly in bacteriophages. Three elements travel together: the reverse transcriptase, a partner protein whose role is unclear, and a long array of DNA repeats spaced at regular intervals. The repeats are the arresting part, because that architecture is the signature of CRISPR, the bacterial defence system that became the basis of modern gene editing.
Early laboratory work adds to the resemblance, indicating that the ART array is transcribed into a set of distinct short RNAs, exactly the arrangement that would allow a system to be programmed to recognise particular sequences. What ART does in nature remains open, and Anthropic has released a preprint and asked outside researchers to propose experiments, framing the discovery as an invitation rather than a finished result.
- hypothesis
- an idea that a scientist tests with experiments
- in parallel
- at the same time, side by side
- comb
- to search something very thoroughly
- flag
- to mark something as worth attention
- architecture
- the overall design or arrangement of a structure
- signature
- a feature that clearly identifies something
- transcribe
- in biology, to copy a DNA sequence into RNA
- framing
- presenting something in a particular way
Level 4 — Advanced
The claim Anthropic advanced on 23 September 2026 is narrower than the headlines it generated and more interesting for being narrow. Roughly 950 instances of its Claude model, given high level direction rather than a protocol, spent 21 hours and 210 million tokens traversing genomic databases and returned with a candidate system that the literature does not appear to describe. The novelty resides in the search strategy as much as in the object found.
Methodologically, this is triage at a scale no human group would attempt. The agents assessed more than 200,000 reverse transcriptases, enzymes whose ability to write RNA back into DNA makes them favourite raw material for mobile genetic elements, promoted some 3,500 loci to candidate status, and compressed that set to twenty for close reading. The pipeline is best understood as a filter tuned for strangeness rather than for confirmation, which is precisely the regime in which exhaustive machine search has an advantage over expert intuition.
The system, designated ART for array associated reverse transcriptases, occurs predominantly in bacteriophages and comprises three co-located elements: the transcriptase, an accessory protein of undetermined function, and an extended array of regularly spaced DNA repeats. That last feature carries an unmistakable resonance, since regular spacers adjacent to an effector enzyme are the structural hallmark of CRISPR, and preliminary bench work indicating that the ART array is expressed as a population of discrete short RNAs strengthens the analogy considerably.
Restraint is nonetheless warranted. A CRISPR-like architecture does not establish a CRISPR-like function, and Anthropic is explicit that what ART accomplishes in its native context is unresolved. By issuing a preprint and soliciting collaborators rather than a claim of mechanism, the company has positioned the work where it properly belongs: as a well characterised anomaly handed to the wider field. The durable question raised is procedural, namely whether large scale agentic search becomes a standard instrument of genomics, in which case the shortage in biology shifts from hypotheses to the bench capacity required to test them.
- protocol
- a fixed set of steps for carrying out an experiment
- traverse
- to travel across or through something
- triage
- sorting items by priority so the most urgent get attention first
- locus
- a specific position on a chromosome or in a genome
- hallmark
- a distinctive feature that identifies something
- effector
- a molecule that carries out an action in a biological system
- anomaly
- something that does not fit the expected pattern
- procedural
- relating to the method or process used, not the result