On September 23, Anthropic published a research post claiming that Claude had found a previously uncharacterized enzyme system in bacteriophages — three parts: a reverse transcriptase, a partner gene beside it, and a long array of evenly spaced DNA repeats that looks like the architecture of a CRISPR array. The company calls it ART. It is, if it holds, a nice piece of machine pattern-matching: the underlying reverse transcriptase had already been described in a jumbo phage in earlier work, and what Claude contributed, per Anthropic, was noticing the repeat array and an accessory protein of unknown function sitting next to it.

As of Wednesday, nobody outside the building had confirmed the finding. That’s the whole ballgame for the skeptics, and the skeptics are right, and it does not matter even slightly.

Because the argument everyone is having — did the chatbot really discover something, or is this a press release with a nucleotide in it — is a category error. The enzyme is not the product. The queue is.

The Reverse Transcriptase Was Already in the Literature

Yes. Point taken. A finding that leans on prior published work and hasn’t been replicated is not a discovery in the sense a bench scientist uses the word. The claim may be right; it may be a confident summary of something a graduate student would have circled in 2023. Anthropic’s own framing is careful: Claude “appears to be the first to notice” the defining features. Notice. Fine.

But notice how little that costs them. A biology claim like this doesn’t need to survive replication to do its job. It needs to survive a news cycle. And in the same week, Anthropic announced that Claude had accelerated more than 30 open-source biomolecular models by roughly 4x on average, opened public beta applications for a new Life Sciences Verification Program, and co-sponsored a protein design competition with Adaptyv Bio covering five hard problems — species cross-reactivity, pH sensitivity, peptide-MHC specificity, and difficult GPCR targets among them.

That is a lot of scaffolding for one enzyme nobody has checked.

Anthropic Bought the Only Instrument That Can Check Anthropic

On September 16, Novo Nordisk announced a collaboration to test Claude Science inside its own R&D workflows. Anthropic already sells to Roche’s Genentech and has an arrangement with Bristol-Myers Squibb. Novartis CEO Vas Narasimhan sits on Anthropic’s board. And the company acquired Coefficient Bio for roughly $400 million in stock, which is how you buy a wet lab without the indignity of building one.

Put those pieces together and you get a genuinely novel structure: the entity making the biological claim owns the laboratory that would test it, the models the testers would use, and the commercial relationships with the firms that would monetize the result.

Here is the uncomfortable part. Replication is a public good with no buyer. Nobody gets tenure for confirming someone else’s enzyme system, nobody’s stock goes up, and no pharma company pays for the privilege of checking a competitor’s vendor. That was always the load-bearing weakness of institutional science — the checking was underfunded and thankless. It was also, until now, done by people with no stake in the answer.

One staff scientist at a mid-sized university structural-biology lab, replying in a public Slack channel for an open-source protein-modeling group, put it more bluntly than I will: “The competition is scored by the people whose models you have to use to enter it. That is not verification. That is a loyalty program.”

Vetted Labs and the Problem of the Vetter

The Life Sciences Verification Program is open, in beta, to “vetted” labs and biotechs. Ask the obvious question: vetted by whom, against what standard, and with what recourse if you’re declined? There is no appeal, no docket, no opinion you can read. There is a form.

I don’t think anyone at Anthropic is twirling a mustache. The company publishes a great deal and hands out model access more freely than its rivals. But the direction of travel is unmistakable. For forty years the independent check on a scientific claim came from a university lab that competed with the claimant for credit, funding, and priority. Now that lab is an applicant. It fills out a form, waits for a beta slot, and agrees to terms it did not write, in order to check work produced by the company that decides whether it gets the slot.

That’s not a conspiracy. It’s an org chart.

The Right Can’t Cheer This and Still Believe in Replication

If you spent the last fifteen years explaining that peer review is captured, that journals are cartels, that grant panels protect their own, then September 2026 should bother you more than an NIH study section ever did. At least a study section leaves a paper trail you can request.

Here there is nothing to request. Private lab. Private models. Gated beta. A partner list that reads like a portfolio.

The reflex on my side of the aisle is to hear “private” and relax. But private gatekeeping is still gatekeeping, and this one has better lawyers than a federal agency and no FOIA officer at all. Meanwhile the reflex on the other side — trust the institution, the science will out — assumes an institution that no longer has the budget, the instruments, or the incentive to do the outing.

ART may well be real. If it is, someone eventually puts it in mammalian cells, and the world gets a better editing tool, and we should all be glad. But the durable artifact of this September is not an enzyme. It’s a turnstile, and Anthropic is standing next to it, holding the clipboard.

Sources