On 23 September Anthropic announced that its model Claude had surfaced a previously uncharacterised enzyme system. The system is called ART and sits in the DNA of bacteriophages, the viruses that infect bacteria.

What makes it interesting is the structure beside the enzyme: a long array of evenly spaced DNA repeats, reminiscent of the arrays in CRISPR systems. In CRISPR, those arrays act as a bank of RNA guides, which is what makes the system programmable.

What Claude did

The work was done under the life sciences programme Anthropic set up this year. The numbers:

  • More than 200,000 reverse transcriptases were gathered from databases. These are enzymes that copy information from RNA into DNA.
  • About 950 Claude agents searched through them for 21 hours, spending roughly 210 million tokens.
  • The search produced 3,500 candidate systems, of which 20 were picked for closer analysis.
  • One agent noticed a CRISPR-like repeat array next to the gene of one of those candidates.

By Anthropic's own account, the model's contribution was not discovering a new enzyme. The enzyme itself had been described before, in a jumbo bacteriophage. What is new is putting the repeat array and a partner protein of unknown function together, and noticing that this may be a distinct system.

What is not known

With stories like this, the point is reading where the claim actually stands:

  • The biological function of ART is unknown. Anthropic says it is under investigation.
  • Early lab work showed the array expresses short RNAs, which could point to CRISPR-like programmability but does not prove it.
  • What the partner protein does is unknown.
  • The work was published as a preprint and has not been peer-reviewed.

So "AI found a new gene-editing tool" is not a correct sentence today. The correct one: the model surfaced a candidate that human researchers found worth examining.

The lab itself

Anthropic opened its own molecular biology lab in the Bay Area in the spring of this year. By the company's account, it operates only at the two lowest biosafety levels and works with no human pathogens. The experiments are run by human scientists, not models; hypotheses the model produces are assessed by researchers first.

That division is one of the clearest illustrations of AI's current role in science. What Claude did was comb through 200,000 samples and flag a pattern a person would likely have missed; the hands holding the test tube are still human. Whether the finding holds will be settled by experiments and peer review.