Back in September I wrote that something genuinely remarkable had happened in science. Anthropic's Claude, running as hundreds of agents over a single day, had spotted a strange CRISPR-like pattern in the DNA of viruses that nobody had noticed. I nearly dropped my coffee. I wrote about Casgevy, the sickle cell medicine I helped launch, and how that whole story began with a pattern of repeats nobody understood.
Then the comments arrived.
A Chief Scientific Officer at a genomics company said the system had already been published and Anthropic had failed to cite it. Another scientist was blunter. He said I'd written cheerleading science journalism, where the "rah rah" leads and the "hold on a minute" gets buried.
My first instinct was to get a bit defensive. My second, better one was to go and find out whether they were right.
So I went away and read everything I could find: the preprint, the coverage in Nature, the New York Times, Wired and STAT, and what working scientists have been saying on Bluesky and X. Here's what I found. It's more interesting than either of us thought.
So, was it already known?
Partly, yes. And to be fair to Anthropic, they said so themselves.
The enzyme at the centre of the system had been described before. Jason Gill's lab at Texas A&M reported it in 2021 while cataloguing three new viruses that infect Staphylococcus. Anthropic's announcement says plainly that it "had been identified in previous studies", and the full preprint cites Gill's paper along with the other relevant work. What they claimed was new were the things sitting next to the enzyme: a long array of evenly spaced repeats and a mystery partner protein. Gill's paper doesn't mention either.
So on the narrow question of published work, the "they didn't cite it" charge mostly doesn't hold. The bigger miss was in the blog post, which is what most people (me included) actually read. It mentioned "previous studies" without naming anyone, and that's where the credit quietly fell out of the story.
But then it gets much more interesting.
The scientist in Copenhagen
Four days after the announcement, Carl Zimmer at the New York Times introduced the world to Mario Rodríguez Mestre, a researcher at the University of Copenhagen. He and his colleagues had been studying these exact enzymes for about four years. They'd even given them a nickname, "jumbotrons", because they live in giant viruses called jumbo phages.
His work hadn't been published yet. It sat in his dissertation and his draft manuscripts.
And he had pasted those drafts into Claude.
"The most important question is not 'Were ARTs already known?'" he told the Times. "They were." Seth Shipman, a respected genome-engineering researcher at Gladstone and UCSF, backed him up: "We've been talking with him about these for years. Many of us know about these systems."
Mestre has been careful to say he has no proof. His worry, as he put it to The Scientist, is "that this information was used to train future versions of the models." Anthropic's response has been firm: "We are not aware of any previously published work describing the ART system we recently found. Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either."
I want to be really fair here, because I use Claude every single day and I'd be lying if I said I had no stake in this. Nobody has shown that Mestre's chats went anywhere near the model that made the find. A group at Stanford posted a closely related find the very same week, a similar repeat array sitting next to a different family of enzymes, which tells you these patterns really are sitting in the public data waiting to be noticed. It's entirely possible Claude found it on its own.
But notice what's happened. The only people who could prove it either way are the company making the claim.
It's happened before. Two weeks before.
On 8 September, OpenAI announced that one of its internal models had cracked a major result on the Navier–Stokes equations, one of the famous Millennium Prize problems in maths. Within hours, NYU mathematician Tristan Buckmaster said he'd been working on the same route with his collaborator Levent Alpöge for a year, putting "all our drafts for the whole of this project" into OpenAI's own Codex.
Buckmaster says OpenAI suggested he write the paper alone, crediting their model, without Alpöge, who happens to work at Anthropic. He refused. OpenAI's own statement included a sentence I keep rereading: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
Two of the biggest AI labs in the world. Two "AI discoveries" in the same month. And in both cases, a human researcher who had been working on the very same problem, inside that company's own product.
That's the story I missed.
Science has always fought about who got there first
None of this is new in one sense. Discovery is famously messy about credit. The sociologist Robert Merton spent a career showing how often big ideas are found by several people at once, and how credit tends to flow to whoever is already famous. He called that the Matthew effect: to those who have, more shall be given.
CRISPR itself is the perfect example. Francisco Mojica worked out what those repeats were for in 2003 and had his paper rejected by Nature, PNAS and others. Years later, the people who turned CRISPR into a gene-editing tool spent the best part of a decade fighting over the patents. Credit in science is never a footnote. It's careers, funding and, sometimes, billions.
And AI has form for "discovering" things that turn out to be in the literature already. Last year OpenAI's people claimed GPT-5 had solved ten unsolved Erdős maths problems, until the mathematician who maintains the list pointed out the model had mostly found existing solutions he hadn't known about. Demis Hassabis's verdict was one word: "embarrassing". To their credit, a DeepMind team later published a careful study finding that 8 of the 13 "open" problems they tackled were already solved somewhere. Many were open "through obscurity rather than difficulty", they wrote, and they warned about the risk of "subconscious plagiarism" by AI.
Subconscious plagiarism has a proper name in psychology: cryptomnesia. A memory comes back to you and feels like your own original thought. George Harrison was found to have done it with "My Sweet Lord", which a judge ruled was the same song as "He's So Fine". The judge didn't think he did it on purpose. It didn't matter. He'd heard it.
What's genuinely new in 2026 is who sits on the other side of the argument. When a rival lab beats you to a result, you can read their paper and check. When the "rival" is the company whose software you used to do your thinking, the evidence (the training data, the chat logs) is all on their side of the wall. You're left trusting a press statement.
The scientific establishment already suspects this is a problem. Since 2023 the US National Institutes of Health has banned reviewers from putting grant applications into AI tools, because there's "no guarantee of where data are being sent, saved, viewed, or used in the future." Yet thousands of researchers paste their own unpublished work into these same tools every day, because the tools are brilliant and the deadline is Friday. Mestre has now said he's stopping.
What I got wrong, and what I'd still defend
I still think the discovery is exciting. Feng Zhang called it "genuinely intriguing", and he should know. Jason Gill, whose own paper missed the repeats, said the models are better at spotting these patterns "than a person staring at it with their eyeballs". Shipman put it well: "The novel thing is how they found it, not what it is."
What I got wrong was the shape of the story. I told it as a lone AI having a eureka moment. The truth looks more like how science has always worked: a crowd of humans had been circling this thing for years, a machine got there fast, and the arguments about who deserves the credit started the moment the press release went out. One Stanford scientist, Le Cong, summed it up in Wired: "The experiments are still in the queue. The PR is already live."
And here's the bit that genuinely gives me hope. The very same technology that may have blurred the credit is also the best tool we've ever built for finding it. These models are astonishing at digging up the obscure 1977 paper or the half-forgotten thesis. That's literally what tripped up the Erdős claims. So if a lab wants to announce an AI discovery, the first agent it runs should be the one asking "who got here before us?", and the answer should be printed right at the top of the announcement, names and all.
I'd love to see Anthropic credit Mestre properly. I'd love the next announcement to lead with the humans who were already on the trail. And I've learned something about my own writing: when the "?!" is that exciting, that's exactly when I should go looking for the "hold on a minute".
To the scientists who pushed back in my comments, thank you. You were right to.
If you want the daily version of this, the AI stories that genuinely matter, that's what Briefly AI is for. Six minutes, every morning, on your podcast app of choice.