Pull to refresh
Logo
What Changed
Following Why Sign Up
Claude AI discovers CRISPR-like enzyme system in phage DNA

Claude AI discovers CRISPR-like enzyme system in phage DNA

New Capabilities

Outside experts say ART is not a gene editor and its RNAs may be templates; Anthropic's 10 reruns missed the array, and no independent lab had validated it as of October 9

October 26th, 2026: Outside experts say ART is not a gene editor

Overview

Updated Oct 27

Anthropic's Claude spent about 21 hours searching more than 200,000 reverse transcriptases, enzymes that copy RNA into DNA. It flagged one sitting beside a repeating DNA array that resembles CRISPR's layout. Lab tests showed the array produces short RNAs, but the partner enzyme, named ART, has no known function.

Anthropic reran the search 10 times and all 10 missed the array. Outside experts now say ART is not a gene-editing tool: it has no DNA-cutting enzyme, and a reverse transcriptase only writes DNA. Stanford's Alex Gao thinks the RNAs are templates rather than CRISPR-style guides, a test that cDNA experiments could check. No independent lab had validated the finding as of October 9.

Why it matters

ART could become a gene-editing tool if programmable, but no independent lab has validated it and missed reruns show AI discovery depends on luck.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

10
Reruns that missed the ART array
Anthropic reran the same agent campaign 10 times; no rerun read the DNA upstream of the enzyme, so all 10 missed the array.
0
Independent lab validations as of October 9
No laboratory outside Anthropic had reproduced the wet-lab work as of October 9, according to an analysis of the preprint.
90%
Array found when placed in context (best model)
In a fixed-input benchmark, four Claude models described the array in at least 90% of attempts when the relevant loci were placed directly in their context.
32%
Array found in tool-using setup (lowest)
Performance fell as low as 32% in a tool-using setup, where models had to work with files and tools rather than direct context.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

CM
Claude (Anthropic AI model)
Autonomous discovery agent that identified ART
Identified ART after a 21-hour autonomous database search
 Dario Amodei
Dario Amodei
Chief executive of Anthropic
Wrote that ART's function and significance are not yet clear, and noted a Stanford team had independently described a different RT system with a non-coding array
 Lucas Harrington
Lucas Harrington
Co-founder of Mammoth Biosciences
Calls the find routine genome mining; says the method is standard practice in CRISPR discovery
 Kevin Blake
Kevin Blake
Microbiologist at Washington University School of Medicine
Skeptical of ART's significance as a gene-editing tool
 Feng Zhang
Feng Zhang
Professor at MIT and the Broad Institute; helped turn CRISPR into a gene-editing tool
Called the work 'genuinely intriguing' and said it merits further investigation
 Aaron Engelhart
Aaron Engelhart
Associate professor at University of Minnesota
Says findings are preliminary and don't demonstrate gene editing

Organizations Involved

Anthropic
Anthropic
AI company
Discovered and announced the ART enzyme system

San Francisco-based AI company that builds the Claude model and led the research program that found ART.
Mammoth Biosciences
Mammoth Biosciences
Biotechnology company
Offered skeptical commentary on the discovery

Uses genome mining to discover CRISPR systems, the same general approach Anthropic employed.

Timeline

September 2026 October 2026

6 events Latest: October 26th, 2026 · 2 weeks ago
Tap a bar to jump to that date

  1. Outside experts say ART is not a gene editor

    Latest Reaction

    Debojyoti Chakraborty (CSIR-IGIB) and Kutubuddin Ali Molla (ICAR-NRRI) say ART lacks the DNA-cutting enzyme gene editing needs. Alex Gao (Stanford) argues its RNAs are likely templates, not guides.

  2. Analysis finds no independent validation of ART

    Analysis

    An analysis of Anthropic's preprint found no independent laboratory validation as of October 9. The registry What AI Found classifies ART as a claimed, collaborative result. A benchmark showed models found the array in 90% of attempts in context, but as low as 32% with tools.

  3. Biologist claims earlier identification of the system

    Dispute

    A biologist says he identified the system in 2022 and that his team studied it for four years without publishing. A 2023 patent filing is also cited. The claim has not been independently checked.

  4. Anthropic discloses 10 reruns missed the ART array

    Disclosure

    Anthropic ran the same agent campaign 10 more times; no rerun read the DNA upstream of the enzyme, so all missed the array. In a controlled test, four models given the sequence directly found the array in at least 90% of attempts.

  5. Anthropic publishes preprint announcing ART enzyme system

    Publication

    Claude agents searched for 21 hours, examined 200,000+ reverse transcriptases, and flagged the ART pattern. Lab tests confirmed the array produces distinct short RNAs.

  6. CRISPR researchers call finding routine genome mining

    Reaction

    Lucas Harrington notes similar systems known since 2008; Kevin Blake says nothing indicates ART is a rival to CRISPR-the-technology.

Scenarios

1

ART confirmed as a functional programmable system

Unlikely Resolves by Q3 2027

Discussed by: Anthropic's research team, whose preprint describes ART's architecture as resembling programmable systems

Lab experiments show the ART reverse transcriptase is enzymatically active, processes array-derived RNA, and performs a programmable operation analogous to CRISPR's spacer bank. That result could make ART a new molecular tool. It would also validate Anthropic's autonomous approach as a genuine discovery engine rather than confirmation of known biology.

2

ART's natural function found, but it's no gene editor

Possible Resolves by Q3 2027

Discussed by: Microbiologists like Kevin Blake, who note ART resembles RT-associated defense systems already catalogued

Researchers determine ART's native role — likely a defense system in the phage-bacteria arms race — and show whether the RT is active. That yields insight into RT biology but no programmable editing capability. ART becomes a data point on a long list of genome-mined systems that never became tools.

3

ART research stalls, function never confirmed

Possible Resolves by Q3 2027

Discussed by: Lucas Harrington of Mammoth Biosciences, who calls the finding routine genome mining

No lab produces evidence the ART reverse transcriptase is active or that its RNA array is processed. The system fades from attention, joining the many CRISPR-like sequences in unstudied bacteria that Blake says remain uncatalogued. The episode matters mainly as a demonstration of AI autonomy, not as a biological breakthrough.

4

ART discovery attributed to chance, AI method shown unreliable

Possible

Discussed by: readus247.com analysis of Anthropic's rerun data

Anthropic's own data shows the agent campaign missed the ART array in 10 of 10 reruns. If the method can't reproduce its own find, the episode becomes evidence of AI's limits rather than its potential, and ART stays a lucky catch with no demonstrated function.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

1987

CRISPR repeats first observed (1987)

Yoshizumi Ishino's lab at Osaka University reported odd repeated DNA sequences in E. coli with no known function. The pattern, later named CRISPR, sat unexplained for years.

Then

The repeats were cataloged as biological curiosities; researchers had no idea they encoded an adaptive immune system.

Now

In 2012, Jennifer Doudna and Emmanuelle Charpentier showed CRISPR-Cas9 could be programmed to edit DNA, a discovery that won the 2020 Nobel Prize in Chemistry.

Why this matters now

ART is a similarly odd repeating pattern with unknown function. The open question is whether it follows CRISPR's path to utility or stays a curiosity.

2008–2026

Genome mining for CRISPR and RT systems (2008–present)

Labs have spent nearly two decades scanning microbial DNA databases for CRISPR-Cas and reverse-transcriptase systems. Companies like Mammoth Biosciences built their businesses on this approach, which has produced many candidate systems.

Then

Most mined systems never become tools; the field treats them as raw material.

Now

The approach became standard practice, which is why Lucas Harrington calls Anthropic's search routine.

Why this matters now

Anthropic's contribution is autonomy: AI agents decided which leads to pursue and flagged ART for lab tests. The mining method itself is not new.

Sources

(23)
+17
Anthropic The Decoder Al Jazeera The Next Web Anadolu Agency AllSci Big News Network Interesting Engineering GIGAZINE Pubt ReadUs247 Mixed News Model Current OpenTools KRDO Smithsonian WKOW AI In Context Pharmatica CellCog India Today Scientific American ANOTHER News