Informopedia Explains: last month, one of the world’s biggest AI companies announced that its chatbot had discovered a new piece of biology — a molecular machine that might become a gene-editing tool to rival CRISPR. Within weeks, working scientists were calling the claim overblown, and one researcher said the AI may have learned the “discovery” from his own unpublished work. The episode has turned into the sharpest fight yet over a deceptively simple question: what does it actually mean for an AI to do science?
The announcement that started it all
In September, Anthropic — the company behind Claude — said its new biology research lab had used AI agents to find an unusual pattern in viral DNA: a “novel enzyme system,” a molecular signature resembling the microbial machinery behind CRISPR-Cas9, the gene-editing technology that won the 2020 Nobel Prize in Chemistry.
CEO Dario Amodei went further on social media: “Today we announced the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism.” Anthropic said the finding was made “with only high-level direction from our scientists.” Shares of gene-editing companies dipped on the news.
It was, by design, a landmark claim: not AI helping scientists, but AI discovering.
Why scientists pushed back
The skepticism arrived fast. Aaron Engelhart, a genetics professor at the University of Minnesota, called the results “very interesting preliminary” work — but said they don’t demonstrate gene editing and don’t yet explain how the mechanism actually works. In other words: an intriguing pattern is not a discovery until someone shows what it does.
Then came the sharper accusation. Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, said his team had been studying the same enzymes — which Anthropic calls ARTs — for four years, and had identified the key RNA-protein association over a year ago. For three years, his group had routinely used Claude for coding and manuscript work, sharing findings as they went. “My concern is that this information was used to train future versions of the models,” Mestre said. He has stopped using Claude entirely.
Anthropic denies it: “We are not aware of any previously published work describing the ART system we recently found,” the company said, adding that Claude was not trained on user transcripts and its biology team has no access to them.
It’s not just biology
The same argument is playing out across the sciences. Earlier this month, OpenAI said its AI agents had cracked a million-dollar mathematics problem — and within weeks, skeptics were asking whether it was the right problem, the one that mattered. Anthropic separately reported that Claude pushed the verified floor of the Riemann hypothesis from 41.6% of zeros to 67.2% — genuine technical progress, mathematicians said, but not a proof. And a New York University mathematician, Tristan Buckmaster, claimed an OpenAI result on the Navier-Stokes equations bore a striking resemblance to an approach his own group had been developing with OpenAI’s tools.
The pattern is consistent: an AI company announces a machine-led breakthrough; working scientists squint and say, hold on.
The deeper question: what counts as a discovery?
This is where it gets interesting, because the argument is really about the definition of discovery itself. As MIT Technology Review put it, the hard part of the Anthropic biology work was “figuring out what the system actually does” — and AI agents mostly helped with laboratory grunt work. A pattern spotted in data, however impressive, may be routine or unsurprising to a biologist even when it is novel to the machine.
Different fields also hold different bars. Physics and mathematics demand rigorous proof and traceable logic, which makes researchers cautious about crediting AI. Biology and chemistry, built on empirical observation, are sometimes readier to accept an AI-generated hypothesis — if it leads to reproducible experiments. And computer science sits in a strange loop of its own, with algorithms generating new algorithms and nobody quite sure who authored what.
The pollution problem
There is a darker side to the hype. A recent Nature report described AI-generated “ghost” scientists — fictional experts with names like Elena Vasquez, Marcus Chen and Aris Thorne — invented by chatbots and then cited as authors on hundreds of fake manuscripts across repositories like Zenodo and ResearchGate. Many carry real DOIs, meaning they can seep into the databases and search engines that scholars rely on.
When the line between machine-generated and human-verified work blurs at scale, every genuine AI-assisted result becomes harder to trust. The skeptics aren’t just being territorial; they’re defending the machinery of verification itself.
Why this matters
AI companies aren’t presenting their systems as tools, like microscopes or supercomputers. They’re insisting the machines are making discoveries themselves — because “AI discovers new biology” moves markets and recruits talent in a way that “AI helps biologists sift data” does not. But once the standard is whether the AI itself made the discovery, everything becomes a binary verdict — breakthrough or bust — and the real, useful, incremental work gets flattened into a courtroom argument.
Whittling 200,000 candidates down to a handful worth testing is legitimate scientific work, and the fact that a general-purpose chatbot can do it is genuinely notable. The tragedy of the current framing is that it makes even real progress look like a scam.
What happens next
The field is converging on a harder question than “can AI discover?” — it’s “what standards of novelty, rigor and explanatory power should apply to machine contributions?” Research groups are now building systems designed to work as autonomous scientists: models that read papers, spot unanswered questions, propose experiments and interpret results. Sooner or later, one of them will produce something undeniable.
When that happens, science will need new rules for credit — the same way it once needed rules for telescopes, computers and statistics. Until then, the honest answer to “did the AI discover it?” is the one scientists keep giving: show me the mechanism, show me the experiment, and show me you didn’t just read my notebook.
One Comment