In September 2026 Anthropic, the company behind the Claude AI models, said its new molecular biology lab had made its first discovery: a previously unknown set of genes in viruses that infect bacteria, with a layout reminiscent of CRISPR[1][2]. The headlines ranged from “AI discovers new gene-editing tool” to sceptical “the science isn't confirmed”. Both contain some truth. This post explains what was found, how, and how to judge a claim like this, which is a skill your biology exams increasingly test.
First, the biology you need
GCSE coreBacteriophages (phages) are viruses that infect bacteria. Bacteria fight back in many ways, and phages fight back against those defences. That arms race has produced many of biology's most useful tools: CRISPR-Cas9 itself began as a bacterial defence system that stores short pieces of viral DNA between repeated sequences, so the cell can recognise the virus next time. You can read how that became a gene-editing tool in our CRISPR post.
A reverse transcriptase is an enzyme that copies RNA back into DNA, running the usual DNA → RNA direction backwards. You may know it from HIV, which uses it to insert its genes into human cells. Reverse transcriptases are found in many bacteria and phages too.
What Claude found
Anthropic called the system ART, short for array-associated reverse transcriptases[2]. It has three parts that sit side by side in phage DNA:
- a reverse transcriptase enzyme;
- a partner gene beside it;
- an array of evenly spaced, repeated DNA sequences.
Each part had been seen before. What had not been described was the three of them working as one unit, and in particular the repeat array, which sits upstream of the enzyme gene, where earlier searches had not looked[2].
How the search narrowed down. Figures from Anthropic's announcement as reported by The Next Web: about 949 agent sessions over 21.5 hours.
How the search worked
A-Level extensionRoughly 950 Claude “agents” (separate AI sessions, each working on part of the job) ran for about 21 hours through public DNA sequence databases[1][2]. They gathered about 200,000 reverse-transcriptase gene clusters, scored 3,564 families of genes found next to them, and wrote detailed reports on 20 candidates. One agent read the DNA just upstream of the enzyme and noticed the repeat array.
Human scientists then went to the lab. In phage data from Staphylococcus, the array was read out as a set of distinct short RNAs that made up as much as 8% of the phage's RNA 15 minutes after infection[2]. That is a strong hint the system matters to the virus. But at the time of the announcement the team had not shown that the enzyme is active, or that it acts on those RNAs[2].
Quick check
What does a reverse transcriptase do?
The caveats that matter
The distinction students often miss
'Found a candidate' is not the same as 'discovered what it does'.
There is a second, more surprising caveat. Anthropic re-ran the same campaign ten more times. None of the reruns read the DNA upstream of the enzyme, and all ten missed the array[2]. In other words, the discovery depended on one agent making one particular choice. Anthropic published this itself, which is good scientific practice: it tells readers how reproducible the method is.
How to judge a “breakthrough” headline
| Question to ask | For the ART announcement |
|---|---|
| Is it peer reviewed? | No: a preprint at announcement |
| What has been measured, not just predicted? | The repeat array is transcribed into short RNAs during infection |
| What is still unknown? | Whether the enzyme is active, and the system's function |
| Is it reproducible? | The computational search found it once in eleven runs |
| What do independent experts say? | Interested and positive, asking for further work |
| Who benefits from the claim? | The company announcing it, so read its claims carefully |
Exam-style question
A company reports that an AI system has identified a new group of genes in bacteriophages. Suggest two reasons why other scientists might be cautious about the claim, and one way the claim could be tested. [3 marks]
Show model answer and mark breakdown
- the work has not been peer reviewed, so the method and data have not been independently checked (1 mark)
- the result was not reproducible: repeated runs of the same search did not find it again (or: the function of the genes has not been shown experimentally) (1 mark)
- test it by showing in the lab that the enzyme is active and what it does, e.g. knock out the genes and see the effect on phage infection, and have other groups repeat the work (1 mark)
Examiner focus: separate the limitation of the evidence (no peer review, no function shown, not reproducible) from the method of testing it (experiment plus independent replication).
Common mistake: saying 'AI can't be trusted'. The point is not who found it but what evidence exists. The same caution applies to a human-led discovery.
Why this matters for students
Whatever ART turns out to do, the story is a neat example of where science is heading: AI that searches data at a scale no person could, combined with human experiments and peer review to decide what is real. Being able to explain both halves is useful for A-Level biology, for computing, and for any university interview. If you want to see how AI itself learns, read our explainer on backpropagation, and for a maths angle, can AI actually do maths?
Want to revise this topic?
Mostak offers online one-to-one GCSE and A-Level tuition focused on understanding, exam technique and application.
Sources and further reading
- Phys.org: Anthropic touts AI-led biology discovery (Sep 2026) — announcement, preprint status, human-run experiments
- The Next Web: Anthropic says Claude found a new enzyme system with CRISPR-like repeats — ART components, search figures, the ten reruns, Feng Zhang quote
- WION: Anthropic says Claude found a new gene-editing enzyme on its own. The science isn't confirmed — sceptical coverage
Last reviewed: 5 October 2026. Figures are typical published values and rules can change. Exam wording varies by specification, so check yours.
Frequently asked questions
What did Anthropic's AI actually discover?
A previously undescribed group of genes in bacteriophage (virus) DNA that Anthropic calls ART, short for array-associated reverse transcriptases. It has three parts that sit together: a reverse transcriptase enzyme, a partner gene next to it, and an array of evenly spaced DNA repeats, a layout that resembles the repeat arrays of CRISPR.
Has the discovery been confirmed?
Not fully. It was announced in late September 2026 in a blog post and a preprint that had not been peer reviewed. Lab work by human scientists showed that the repeat array is read out as a set of short RNAs, but the team had not yet shown the enzyme is active or what the system does.
Did the AI do the experiments?
No. Claude agents searched and analysed DNA databases and wrote up candidates; according to Anthropic, all laboratory experiments were carried out by human scientists.
Does this mean AI can now do science on its own?
It shows AI can usefully search enormous datasets and notice patterns people missed. But the same search was repeated ten times and none of the reruns found ART again, and the biological meaning still depends on experiments and peer review. It is a powerful assistant, not a replacement for the scientific method.
Reader ratings & reviews
Rate this article across a few different things — your rating helps other readers, and helps us improve.
Was this article helpful?
Tell us what you think — a correction, a question it left unanswered, or a topic you'd like covered next.
Prefer To Talk? Book a free video call at a time that suits you: 15 minutes for a quick question, 30 for anything else. Times show in your own time zone.
About the author
Sudershan Soni
Founder & Lead Tutor at Mostak Services — an MSc-qualified Mathematics, Science, Computer Science & STEM tutor with 20+ years of professional experience, teaching students from 11+ and GCSE to A-Level and beyond, online worldwide.
Read full profile