Anthropic's Claude AI discovers new 'CRISPR-like' biological system

On September 24, 2026, Anthropic announced that its Claude artificial intelligence system identified a previously unknown biological system in the DNA of viruses that infect bacteria. The system is called array-associated reverse transcriptases (ART). This was the first press release from Anthropic's newly established life sciences lab, and it emphasized its similarity to CRISPR—CRISPR is a natural bacterial defense system that scientists now use as a gene-editing technology.

But the claim of "discovering a new CRISPR" is premature. ART comprises three key elements: a reverse transcriptase that copies RNA into DNA, a partner gene located beside that enzyme, and a long array of repeated DNA sequences. The reverse transcriptase itself was already known; what is new is the recognition that it may form part of this larger system.

CRISPR has a similar structure, which is the basis for the two being mentioned together. Anthropic's lab experiments also found that the ART array behaves similarly to CRISPR in some respects. However, the key property of CRISPR is that it can be programmed to target and edit specific DNA sequences, whereas whether ART has the same capability is unclear.

Therefore, at present one can only say that ART is architecturally similar to CRISPR; there is no evidence that it is functionally similar to CRISPR. Whether ART itself matters remains an open question.

Starting from a general research task, the Claude AI agent searched more than 200,000 reverse transcriptases, identified 3,500 candidate systems, and selected 20 for detailed analysis.

During the process, an AI agent followed an accidental observation: it noticed a repeated array pattern, compared it with known systems, identified a potentially interesting result, searched the existing literature, and then flagged ART to human researchers.

The individual techniques involved are not a fundamentally new method. Mature tools already exist for finding genes related to one another, clustering sequences, detecting motifs and repeats in sequences, examining which genes co-occur, and computing how RNA molecules fold in three dimensions. A purpose-built traditional bioinformatics pipeline could very likely detect the same patterns.

The research provides no comparison, so it is impossible to judge whether ART is an obvious result of such a pipeline or a hard-to-detect signal among thousands of candidates. The press release claims that this kind of analysis could take researchers "weeks to months of work," but this claim comes from the press release itself and is not an independently verified conclusion.

The important advance lies in autonomy: Claude was given a fairly broad scientific goal, then chose on its own which analyses to perform, which leads to follow, and deemed a certain result worthy of experimental validation. The analysis itself does not exceed the capabilities of existing methods; what makes this work important is another point—AI is beginning to participate in deciding where to point existing tools and which questions are worth asking.

It is currently unclear how many of the 3,500 candidate systems, or of the 20 selected, represent genuinely novel or interesting biology. Finding one interesting result after exploring thousands of possibilities is not the same as proving that the system can reliably identify discoveries.

Claude requires human researchers to conduct experiments to confirm its suggestions. Anthropic CEO Dario Amodei posted on X that it might eventually even be possible for Claude itself to safely perform experiments by autonomously controlling laboratory equipment. This claim comes from his personal post and is a source claim; whether AI can autonomously control laboratory equipment to perform experiments, and when that might be achieved, remains uncertain.

If this approach can be proven to reliably make discoveries, it could become an important application of AI in biology. Modern biological datasets contain more sequences than researchers can examine one by one. AI agents could become another tool for researchers to manage existing tools: conducting large-scale exploratory searches, linking observations to the scientific literature, generating hypotheses, and prioritizing candidates for scientists.

Whether ART itself matters remains an open question, but the value of this research lies in the fact that it hints at a potentially important shift in how AI contributes to science: AI may increasingly help decide where existing tools should be pointed and which questions are worth asking.