Math Preprints Hit a Stunning 25% AI Usage—What This Explosive Surge Means for the Future of Innovation and Investment!
You ever stumble on one of those stats that just knocks the wind outta you? In March 2026, less than 5% of math papers on arXiv mentioned using AI. Fast forward just five months, and boom—over 24% now openly admit AI’s in the game. That’s no slow climb. It’s a rocket sprint that’d make any Silicon Valley startup green with envy. What’s fueling this meteoric rise? A new preprint titled “The Gold Rush in AI4Math: Where Are We Now?” dives deep into 32,944 math submissions, uncovering a seismic shift where AI’s not just proofreading but rolling up its digital sleeves to tackle real math problems. The math world, traditionally slow to change, is now buzzing with a voluntary AI revolution you didn’t see coming. Are we witnessing the dawn of a new era—where human intellect teams up with tireless AI collaborators to smash open problems faster than ever before? Hold on tight, this ride’s just getting started. LEARN MORE

In March, fewer than 5% of math papers posted to arXiv disclosed any use of artificial intelligence. By August 20, that number had crossed 24%. That’s not a gentle upward trend. That’s a vertical line on a chart that would make a growth-stage startup jealous.
The data comes from a preprint titled “The Gold Rush in AI4Math: Where Are We Now?” authored by Jiashun Jin, Zheng Tracy Ke, and Bingcheng Sui, which analyzed 32,944 mathematics submissions to arXiv between March 1 and August 20, 2026. The findings paint a picture of a discipline in the middle of a rapid, and largely voluntary, transformation.
The numbers behind the shift
The headline stat, 24.14% of math preprints acknowledging AI by August, is striking on its own. But the more telling metric is what the researchers call “substantive AI contributions,” meaning cases where AI didn’t just proofread or format but actually contributed to the mathematical work itself.
That figure jumped from 1.39% in March to 14.09% by August 20. In other words, roughly one in seven math papers now features AI doing real intellectual heavy lifting.
Out of the 32,944 submissions analyzed, 3,575 confirmed some form of AI usage. Of those, 1,712 involved what the researchers classified as substantial contributions. The majority of that substantive work centered on proof construction.
The study also tracked which AI tools researchers are reaching for. OpenAI products appeared in over 60% of papers with confirmed AI use. Anthropic’s tools came in second.
Not all math is created equal
AI adoption isn’t spreading uniformly across the discipline. Combinatorics leads in sheer volume of substantial AI use. Metric Geometry, meanwhile, posted the highest rate of AI adoption relative to its size.
Geographically, the US and China dominate. Together they account for roughly two-thirds of weighted author counts in AI-assisted math papers.
Solving open problems, apparently
Perhaps the most provocative finding in the study involves open problems. The researchers identified 717 open-problem records associated with papers making substantive AI contributions. Of those, 71% were reported as fully resolved by their authors.
Preprints on arXiv are not peer-reviewed, and the history of mathematics is littered with claimed proofs that didn’t survive scrutiny. But even accounting for some percentage of premature victory laps, the volume of claimed resolutions suggests AI is enabling mathematicians to attack problems they might not have attempted otherwise.
The dominant mode of contribution is proof construction. AI isn’t replacing the mathematician’s intuition about which problems matter or which strategies might work. It’s acting more like an extremely fast, tireless collaborator that can explore proof paths and verify logical steps at a scale no human can match.
What this means for the field
The study’s authors advocate for clear disclosure of AI usage, which is notable because the current 24% figure only counts papers that voluntarily acknowledge it. The real number of AI-assisted submissions is almost certainly higher.
With OpenAI tools dominating over 60% of confirmed AI-assisted papers, the choice of AI platform is becoming a strategic decision for researchers. Anthropic holds second position, but the gap appears significant. For AI companies, becoming the default tool for mathematical research isn’t just a prestige play. It’s a feedback loop: the more researchers use your system, the more training signal you get on hard reasoning tasks, which makes your system better at hard reasoning tasks.
The concentration of output in the US and China also has implications for global mathematical research. Countries and institutions without access to frontier AI models risk falling behind not because their mathematicians are less talented, but because they’re working without the same tools.



Post Comment