Did Anthropic’s A.I. Really Make a Scientific Discovery on Its Own?, tell me about this news in (https://www.nytimes.com/2026/09/27/science/anthropic-biology-enzyme-mestre.html)

The article examines an announcement by the artificial intelligence lab Anthropic regarding a discovery made by its AI model, Claude, in the field of molecular biology.

Anadolu Ajansı

The Core Story: Anthropic revealed that it established a dedicated life sciences research team and physical laboratory. The company tasked Claude agents with searching public genomic databases containing roughly 1.9 billion protein clusters to locate uncharacterized families of reverse transcriptases (enzymes that synthesize DNA from an RNA template).

Anthropic+ 1

Working through an autonomous multi-agent campaign over approximately 21 hoursconsuming over 210 million tokens across hundreds of sub-tasksClaude combed through the data, flagged candidates, and examined surrounding non-coding DNA. One agent noticed an overlooked sequence beside a viral enzyme and remarked on a repeating DNA structure. Anthropic designated this newly identified setup an “array-associated reverse transcriptase” (ART), primarily found in bacteriophages (viruses that infect bacteria). The architecture features the enzyme, an uncharacterized accessory protein, and a series of evenly spaced DNA repeats reminiscent of CRISPR arrays.

Anadolu Ajansı+ 2

Did the A.I. Really Make the Discovery “On Its Own”? The reporting addresses both what the AI accomplished and where the boundaries lie:

1. What Claude Did Autonomously: Human scientists did not hand-pick the candidate or tell Claude to look for repeat sequences. The AI planned its search tasks, decided to inspect raw flanking DNA, noticed the repetitive patterns that previous human researchers had overlooked in published phage genomes, searched existing scientific literature to confirm novelty, and wrote up candidate reports for human review.

   TNW+ 1

2. What Claude Did Not Do:

• Physical Experimentation: Claude cannot perform wet-lab experiments. Human biologists inside Anthropic’s laboratory conducted the biochemical assays to confirm that the DNA array is actively transcribed into short RNAs.

  Anthropic

• Independent Scope: The AI operated strictly within parameters, datasets, and high-level directives designed and initialized by human researchers.

  Anthropic

• Functional Proof: The biological function of ART remains unknown. It is not yet proven whether the enzyme is active or if it functions like CRISPR or another programmable genetic tool.

  Anthropic+ 1

• Consistency Challenges: Anthropic noted that repeating the automated campaign did not always produce the same result, as autonomous agents explore different paths across vast datasets.

  TNW

Outside experts, such as CRISPR pioneer Feng Zhang, described the finding as intriguing and worthy of further study. Ultimately, the reporting frames the breakthrough not as an artificial intelligence replacing human scientists, but as an advanced form of computational hypothesis generationaccelerating genomic data mining that typically takes human researchers months into a matter of hours, while leaving physical validation firmly in human hands.

TNW+ 1