Mathematics has always moved slowly. A single proof can take years to write and years more to check. This month, an AI company dropped that timeline into the middle of a very public argument.
On October 6, OpenAI uploaded 722 manuscripts to a GitHub repository, claiming solutions to a large number of open problems. The company grouped them into 372 families of results, and the papers came from an internal model that the public cannot use yet. Mathematicians are still working out how much of it holds up.
How we got here
This did not start in October. In May, OpenAI said an internal model had disproved a long-held belief about the unit distance problem, an 80-year-old question in discrete geometry, and that outside mathematicians had checked the proof. Fields medalist Tim Gowers called it “a milestone in AI mathematics” in a companion paper.
Things got messier in the fall. In September, OpenAI claimed a solution to a version of the Navier–Stokes problem, and a mathematician said hours earlier that his team’s recent work had leaked to OpenAI. On September 21, the company said the same internal model had resolved more than 100 long-standing problems and set up a nine-member advisory group on mathematics and AI. The October release was the biggest step so far.
It also followed an earlier stumble. In 2025, an OpenAI executive said GPT-5 had solved ten Erdős problems. It turned out the model had only found existing solutions in the research literature, and the post was deleted. That history is one reason many experts are cautious this time.
What was actually released
The papers span a wide range. According to the New York Times, as summarized by Daily Nous, they cover algebra, number theory, theoretical computer science, mathematical logic and topology. The same report says the average result took only about three hours of computing.
The count is also a little fuzzy. OpenAI says 372 families of results, while the New York Times counted 377 findings. Either way, it is far more than any human team could produce in a few weeks.
The first problems appeared within a day
Less than 24 hours after the release, OpenAI withdrew three manuscripts because of a sign error. It also revised 14 others with proof repairs and corrected statements. A single sign mistake can wreck a proof, and in this case two other papers depended on the same faulty construction.
Reactions were mixed. MIT’s Andrew Sutherland called the quick withdrawal “the responsible thing to do.” Cornell’s Alex Townsend said he expects more errors will turn up and that OpenAI should have shared the machine-verified papers first.
Why many mathematicians are uneasy
The strongest complaints are about how the work was released, not whether the mathematics is interesting.
One group, the Association for Human Mathematics, said releasing over 700 files at once is not scholarship but a show of power. Critics also point out that checking a proof takes real human time. Dumping hundreds of papers at once passes that cost to unpaid outside experts.
Readability is another issue. Physicist Edward Witten, who has been studying a cluster of the papers, said the work is often impressive but the writing is a mess. In his view, a human who understands the results could present them far more clearly. Columbia’s Michael Harris was blunter. He said he would not be reading the papers.
Verification is a concern too. Kevin Buzzard of Imperial College, a leading supporter of computer-checked proofs, said that in his own area only around six results stood out and few if any had been formally verified. Meanwhile, the Clay Mathematics Institute lists one of the famous problems as “active,” not solved.
Not everyone is against it
Some experts see real value. Daniel Litt of the University of Toronto argued that if people want answers to these questions, there is no good reason to ask the company to keep them secret. Computer scientist Stephen Wolfram said at a National Museum of Mathematics event that you can discover new math easily with these tools.
There is a technical reason math is a good test for AI. Like programming, it gives AI a way to tell when it is right. A proof either holds together or it does not.
OpenAI says it is listening to critics. Its advisory group suggested papers should be understandable to humans, should include the chain of thought behind the result, and should credit earlier work that first introduced the key ideas. The group also said that mathematicians will have to judge whether OpenAI actually followed those voluntary guidelines.
What this means for the future of math
Three changes look likely.
Checking will matter more than creating. If AI can produce proofs quickly, the slow part becomes verification. Expect more demand for tools that check proofs automatically and for experts who can review them.
Training will change. Axios reports that mathematicians are already asking what these discoveries mean for graduate work and for training the next generation of theorists. If a model can settle a research-level question in hours, students may need to focus more on judgment, taste and explaining ideas.
Humans are not out of the picture. One commentator argued the technology will be transformative, but it does not end math or remove the need for human mathematicians. Proofs only become knowledge once people understand them and build on them.
What to watch next
The main question is simple. How many of these 722 papers are correct, and how many contain genuinely new ideas? Answering that will take months of unglamorous work by volunteers, journals and verification groups. Until then, the safest way to read the news is that the results are promising but not yet proven.
For students, teachers and anyone curious about math, the lesson is practical. AI is now a serious player in research, but trust in mathematics still comes from careful human checking, and that has not changed.
