Early on September 24 Beijing time, Anthropic announced a new life sciences research group and laboratory, sharing early results: with only high-level direction from its scientists, Claude autonomously searched massive genomic databases and discovered a previously uncharacterized enzyme system — named ART (Array-associated Reverse Transcriptases) — with CRISPR-like repeated DNA structures. Anthropic's official blog says the team focuses on fundamental biology research with Claude: identifying uncharacterized protein families in DNA datasets, generating hypotheses at scale, and testing them through lab experiments.
[1][2]The nut graf: the significance is that Claude's role shifted from assistant to discoverer — the scientist did not ask it a question; it found the object worth studying inside large-scale genomic data and handed it back to humans for validation. Only high-level direction was the most important phrase in the announcement, and it defines a new division of labor in scientific discovery. In Anthropic's own trajectory this is not a one-off demo: the new lab means Claude's biology use moves from an API tool to internal research infrastructure, following the reported April acquisition of biotech-AI startup Coefficient Bio (media put it at roughly $400 million, all stock) — the line has been built with consecutive moves.
On the result itself, ART remains an early-stage finding: it carries CRISPR-like repeated structures and may represent a previously unknown biological mechanism, but enzyme function, editability, and experimental verification all await peer review and independent replication. Anthropic's blog discloses the discovery process and structural features, not functional validation data. That is precisely the boundary this article should underline: AI can dramatically raise the speed and scale of candidate discovery, but between candidate and confirmed mechanism lie long wet-lab experiments; reading AI found it as AI proved it would overstate the current stage.
Competitive context: OpenAI partners with Ginkgo Bioworks and Retro Biosciences, Anthropic bets on an internal lab plus acquisition, Meta collaborates with institutions on frontier research — the route differences among frontier labs in biology are now visible. The methodological lesson matters more: using Claude to generate hypotheses at scale, concentrating human scientists' judgment on validation and decisions, essentially moves the AI-builds-AI logic into scientific discovery. Yesterday's RSI debate was about models building models; this story is about models building knowledge. Both share one question: when AI takes on discovery work, how do humans build an independent verification system for what AI asserts.
A practical reading of the announcement's framing: Anthropic deliberately scoped the claim to discovery with high-level direction, not to autonomous full-loop science. That scoping matters because it is testable — independent groups can re-run the search over the same public genomic databases and ask whether the same candidates surface, whether ART's repeat structure is genuinely novel, and whether the enzyme family holds up in bench experiments. Until then, the scientific payload of this news is the pipeline (hypothesis generation at scale over uncharacterized protein families) rather than the molecule itself. For readers who follow frontier labs, the durable signal is organizational: Anthropic is now running a wet-lab and a research group whose job is to test Claude's biological hypotheses, which means the cost of a wrong AI guess is now paid by Anthropic's own bench — a governance choice other labs will be watching when they staff their own biology efforts.
[1][2]