Math changed its mind in five months.
That’s the short version of a finding that’s been bouncing around my feeds all week. Researchers pulled every arXiv submission in the Mathematics category between March 1 and August 20, 2026 — 32,944 papers — and counted how many disclosed AI use. In March it was 4.75%. By August it was 24.14%. Roughly a fivefold jump in under half a year.
Numbers like that usually mean a tool got cheap, or a tool got good. Either way, as someone who spends most days wiring up bots and arguing with their output, I find the second-order detail more interesting than the headline: of those disclosing submissions, 1,712 involved at least one substantive mathematical contribution from the model, and 1,225 reported AI involvement in proof construction.
Proof construction is the part that matters
Autocomplete for LaTeX is not news. Summarizing a paper is not news. A model participating in proof construction is a different category of work, because a proof is the one artifact in all of software and science that has no tolerance for a plausible-sounding guess. Code can be wrong and still run. A proof that’s wrong is just not a proof.
So the fact that over twelve hundred papers report AI in that loop tells me something about workflow, not just capability. Mathematicians are not famous for trusting things. If they’re letting a model into the proof step and then putting their name on the result, they’ve built some verification habit around it — reading every line, checking the lemma chain, running it past a formal checker, or just rejecting nine suggestions for every one they keep.
That’s the pattern I keep landing on when building anything useful with these models. The generator is the easy half. The checker is the product.
What bot builders should steal from this
You don’t need to be working on open conjectures to borrow the shape of this workflow. A few things I’d pull directly into agent design:
- Separate proposal from acceptance. The component that generates a candidate answer should never be the component that approves it. Mathematicians get this for free from the culture. Your bot needs it designed in.
- Prefer domains with cheap verification. Math has formal proof checkers. Code has tests and type systems. If your task has a deterministic oracle available, wire it in before you tune a single prompt.
- Log the assist. Those disclosure percentages only exist because people wrote down when a model helped. Do the same in your own systems — which outputs came from the model, which were human-edited, which were rejected. That log is your eval set later.
- Expect rejection rates, not success rates. A tool that’s wrong 80% of the time but cheap to check is still a solid tool. A tool that’s wrong 5% of the time with no way to tell which 5% is a liability.
The claims I’m not ready to repeat
There’s a louder story running alongside the arXiv data. OpenAI published a post titled “Sharing AI progress in mathematics,” Scientific American covered a release of hundreds more math results into “a field already in shock,” and a new internal model from an unnamed company is reported to have produced a broad range of new mathematical results — including solutions to the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis.
I want to be careful here, because the sources don’t settle what those advancements actually consist of. “Progress toward the Riemann hypothesis” covers an enormous range of meanings, from a genuinely new technique to a tidy reformulation of something known. The specifics aren’t established, and I’m not going to pretend otherwise to make a better post.
What I’ll say instead: the gap between a result being announced and a result being understood is where most of the actual work lives. Mathematician Daniel Litt described a workshop whose premise — taken as a starting point rather than a debate — was that AI will become reliably superhuman at mathematics. That’s a premise worth sitting with, and it’s notable that it’s now a premise rather than a provocation. But premises aren’t proofs, and announcements aren’t peer review.
Why I think the arXiv number is the real story
Big-name conjecture claims are hard to evaluate from the outside. A disclosure rate across 32,944 papers is not. It’s a measurement of behavior, aggregated, hard to spin. It says a large community of extremely skeptical specialists independently decided these tools were worth their time, and said so in print.
Adoption curves like that tend to show up in software a year or two before they show up in a tool you’d recognize. If proof construction is already in play, the agent patterns we build next should assume a model can contribute real reasoning and that you still need to check every word of it. Those two facts live together comfortably. Design for both.
🕒 Published: