On September 23, 2026, Anthropic presented the initial findings from its AI-assisted biology lab. The company claims that a set of Claude agents identified, in genomic data, a system combining a reverse transcriptase to DNA repeats that are reminiscent of certain features of CRISPR. The result is interesting, but its biological function, its exact novelty, and its usefulness remain to be established: What does this experiment actually demonstrate about AI’s ability to accelerate scientific discovery?
What Anthropic Actually Announced
In the spring of 2026, Anthropic established a life sciences research team and an experimental laboratory in the San Francisco Bay Area. Its goal is to use Claude to explore biological databases, generate hypotheses, and then test those hypotheses through laboratory experiments.[1]
For its first publicly announced program, the company asked Claude to search for unusual examples of reverse transcriptases. These enzymes copy RNA into DNA and play a role in several biological mechanisms. The researchers identified a complex comprising a reverse transcriptase, a long series of noncoding DNA repeats, and a neighboring protein whose function remains unknown. Anthropic named this complex ART, short for array-associated reverse transcriptases.[1][2]
The system has been observed in bacteriophages, viruses that infect bacteria. The comparison with CRISPR stems primarily from the repetitive organization of the sequences and the presence of an associated enzymatic system. Anthropic notes, however, that it does not know the primary function of ART. It is therefore not established that it is a gene-editing mechanism, nor that it could become a medical tool.
What’s Really New
The reverse transcriptase identified in the phage in question was not entirely unknown. It had already appeared in a comparative genomics study published in 2021 on three giant bacteriophages infecting Staphylococcus aureus. [7] The contribution attributed to Claude consists rather in having linked this enzyme to a larger structure, including DNA repeats and an accessory gene.
The potential scientific novelty therefore lies in the characterization of a genetic system that has been little studied to date. The methodological novelty seems more clear-cut: general-purpose agents—capable of writing code, querying databases, and reorienting their search—have explored a data space that a human team would have found difficult to examine exhaustively within the same timeframe.
This distinction is essential. Discovering that an element exists, proposing a function for it, demonstrating its mechanism, and transforming that mechanism into a tool are four distinct steps. At this stage, ART is primarily a promising biological subject for study.
How Claude Contributed to the Scientific Work
The preprint describes an analysis of approximately 1.9 billion protein clusters. Nearly 950 Claude agents worked for just over 21 hours and consumed approximately 210 million tokens. They first identified more than 200,000 candidate reverse transcriptases, then gradually narrowed down this set to isolate unusual families.[1][2][3]
The agents did more than just summarize articles. They performed bioinformatics analyses, compared sequences, examined the genomic neighborhood of genes, and generated research leads. One agent ultimately reported the presence of a regular DNA repeat near a particular family of reverse transcriptases.
The term “autonomous discovery” should therefore be used with caution. Autonomy is concentrated in computational exploration. Humans defined the objective, built the environment, conducted the experiments, and decided how the results should be interpreted and communicated.[4]
Why This Method Is Important for Biology
Contemporary biology generates more sequences, structures, and measurements than a research community can examine manually. The challenge is no longer simply to generate data, but to identify, within these vast amounts of data, the rare patterns that warrant experimental investigation.
AI agents can help broaden this scientific focus. They can scan multiple databases, test groupings, document anomalies, and suggest candidates for further examination. In the case of ART, Claude’s greatest value lies in its ability to filter through a vast space and identify an interpretable signal.[5]
This capability does not automatically extend to all fields of science. It is particularly well-suited to fields rich in digitized data and analytical tools. It is less readily applicable to questions that require long-term observations, costly experiments, rare samples, or contextual knowledge that is difficult to formalize.
The figure of 21 hours must also be considered in light of the cost of calculating it. Two hundred ten million tokens, nearly a thousand agents, and the involvement of a specialized team represent significant resources. Without a comparison to a conventional method, it is impossible to translate this timeframe into a net gain in productivity, cost savings, or quality.
What the results actually allow us to conclude
The preprint reports that ART-associated repeats are expressed during phage infection and appear as distinct RNA units. This finding suggests the existence of an organized functional system, but does not yet reveal its precise biological role.[2]
The paper is a preprint and has not yet undergone peer review. Comments gathered by Nature and WIRED acknowledge the method’s merit, while noting that the significance of the system and its potential role in gene editing remain to be demonstrated.[3][5]
The experiment therefore supports the conclusion that AI agents can quickly identify a promising genetic sequence within a massive dataset. However, it does not yet support the conclusion that Claude has discovered a new CRISPR-like system or produced a therapeutic breakthrough.
Reproducibility, Attribution, and Biosecurity
Agent-assisted research must provide sufficient documentation of its process. The model version, prompts, databases queried, code executed, selection criteria, human interventions, and negative results are necessary to enable evaluation and reproducibility. A final article that published only the selected candidate would obscure the fact that thousands of other leads were ruled out.
The origin of ideas has also become an issue. Researchers have questioned the degree of novelty in the results and the source of certain similarities with previous or unpublished work. There is no public evidence to suggest that Anthropic used this work. The company asserts that Claude is not trained on its users’ conversations and that its biology team does not have access to them.[5][6] The controversy nevertheless highlights the need for attribution and traceability mechanisms tailored to AI-assisted research.
In terms of biosafety, Anthropic states that its laboratory operates at BSL-1 and BSL-2 levels, without handling pathogens capable of infecting humans, and that the experiments are conducted by scientists.[4] These limitations reduce certain immediate risks. However, they do not fully address the issue of dual-use, especially if future systems are capable of designing experiments or directly controlling equipment.
Scientific responsibility remains both human and institutional. A model can provide an analysis, but it cannot guarantee the integrity of the data, approve a protocol, take responsibility for an error, or be held accountable for excessive communication.
What to Watch for Now
The first step will be peer review, followed by independent replication. Researchers will need to determine the substrate of ART, its mechanism, its role in bacteriophages, and whether it can be programmed.
We will also need to compare the system to more traditional methods. How many relevant candidates does it generate? What is its false-positive rate? How much human labor does it save, taking into account the necessary verification, calculations, and experiments? These metrics will allow us to assess the actual acceleration rather than just the execution speed.
Finally, the community will need to establish publication standards for agent-assisted discoveries: activity logs accessible to reviewers, disclosure of human interventions, data sources, and attribution rules. In 2026, the NIH further strengthened its initiatives focused on the replication and reproducibility of biomedical research.[8]
The ART case does not yet demonstrate that AI can replace scientific reasoning and experimentation. It shows something more specific: when a problem can be formulated, implemented, and documented, agents can significantly expand the scope that researchers are able to explore. The next step will not only be to find solutions faster, but to prove them more effectively.
To learn more
To learn more about the connections between artificial intelligence, biological data, health, and the scientific method, check out these analyses from the aivancity blog.
Sources
[1] Anthropic. (September 23, 2026). Claude discovers a novel enzyme system with CRISPR-like repeats. View source
[2] Yoon, P. H. et al. (2026). Autonomous AI agents discover reverse transcriptases with tandem repeat arrays. Preprint. View the preprint
[3] Nature. (September 25, 2026). Anthropic’s AI biolab finds ‘CRISPR-like’ DNA in viruses. What’s next? Read the article
[4] TechCrunch. (September 23, 2026). Anthropic says its biology lab has already made a major discovery. Read the article
[5] WIRED. (September 29, 2026). Anthropic Says It Discovered a CRISPR-Like System. Now What? Read the article
[6] The Atlantic. (October 4, 2026). AI’s Real Gift to Science. Read the article
[7] Korn, A. M. et al. (2021). Comparative Genomics of Three Novel Jumbo Bacteriophages Infecting Staphylococcus aureus. Journal of Virology. View the study
[8] National Institutes of Health. (2026). Strengthening Replication and Reproducibility of NIH-funded Research. View source