From 4.75% to 24.14% in five months. That’s the share of arXiv math submissions disclosing AI use between March 1 and August 20, 2026, drawn from 32,944 papers in the Mathematics category. One in twenty became one in four over a single summer. I review AI tools for a living and I have rarely seen an adoption curve that steep in a field this conservative.
Mathematics is supposed to be the slow discipline. Proofs take years. Referees take longer. And yet the disclosure rate quintupled in about 170 days.
What the numbers actually say
Let me be careful here, because the gap between what was measured and what people will claim it means is about to get very wide.
Of those disclosing submissions, 1,712 involved at least one substantive mathematical contribution from AI. And 1,225 papers reported AI involvement in proof construction. Proof construction. Not formatting, not literature search, not LaTeX cleanup. The part where the actual mathematics happens.
Separately, an unnamed company’s internal model has reportedly produced a broad range of new mathematical results, including solutions to the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis. The sources do not provide a definitive answer on the exact nature or specifics of those advancements.
That last sentence is the one you should sit with. Not because it’s a gotcha, but because it’s the whole review. We have a headline-grade claim attached to the most famous open problem in mathematics, and we do not have the specifics in hand.
Why I’m not calling this settled
“Progress toward the Riemann hypothesis” is a phrase that can mean almost anything. It can mean a genuinely new partial result that specialists will build on for a decade. It can also mean a tightened bound on a lemma inside a subfield of a subfield that was already being tightened by three graduate students. Both are real progress. They are not the same news story.
The four-dimensional Kakeya claim is more legible, because Kakeya is a specific conjecture with a specific statement. But “solution” still needs the thing every solution needs: a proof that other humans have read, checked, and failed to break.
So my honest read on the internal-model claims is: interesting, unverified, and strategically timed. Announcements that arrive without the artifacts attached are marketing until proven otherwise. That’s not cynicism about the underlying capability. It’s the same standard mathematics applies to everyone, which is sort of the point of mathematics.
The adoption curve is the real story
The arXiv numbers deserve more attention than the conjecture headlines, and they’re getting less. Here’s why they matter more:
- They’re measured, not announced. 32,944 submissions, a defined window, a countable disclosure rate.
- They reflect behavior across an entire field, not the output of one lab’s unreleased system.
- They cover the part of the workflow that actually changes careers: 1,225 papers with AI in proof construction.
And one more thing worth sitting with. 24.14% is the disclosure rate, not the usage rate. Disclosure requires a researcher to decide that AI involvement was significant enough to mention, and to be comfortable mentioning it. Those are social decisions as much as factual ones. The true usage number is almost certainly higher than the disclosed one. I can’t tell you by how much, and neither can anyone else right now.
What this means if you’re using these tools
A field in which a quarter of submissions disclose AI involvement is a field where norms are being written in real time, by nobody in particular. There was no vote. There was no standards body. Individual mathematicians made individual calls about what to use and what to admit to, and the aggregate became policy.
That’s how it usually goes with tooling, and it usually works out fine. Mathematics has a specific problem, though: the entire discipline runs on verification. A proof is a social object that becomes true when enough qualified people confirm it. If AI systems generate candidate results faster than humans can check them, the bottleneck moves from production to verification, and the field’s quality control starts operating under load it was never designed for.
1,225 proof-construction papers in under six months is a load. Nobody has told me who’s checking them.
My take
Treat the adoption statistics as solid and the conjecture claims as pending. Use these tools, disclose when you do, and apply the same skepticism to an AI-generated lemma that you’d apply to one from a stranger on the internet, because functionally that’s what it is.
The quintupling is real and it’s the fastest methodological shift I’ve watched a mathematical discipline undergo. The Riemann headlines need a proof I can read. Until I can, they’re a press release with good branding.
🕒 Published: