I have a PhD in AI, and this month I share the grief of a great mathematician.
Hugo Duminil-Copin won the Fields Medal in 2022. On 6 October 2026, OpenAI posted 722 AI-generated math manuscripts on GitHub. Two days later, during a video call hosted by the French Mathematical Society, he said the release had “annihilated” his area of research, and that one night he had cried hard enough to wake up his daughter (OfficeChai, IBTimes).
Most of the debate since then asks whether AI can do mathematics. I think the more useful question is a different one: who will check, understand and explain what the machines produce? That question applies to every expert, and it is where experts survive.
TL;DR
- AI now produces research results at industrial scale. Mathematics fell first because, for a language model, mathematics is a language. Law, accounting and research come next.
- The results are only partly correct. About half of OpenAI’s results were released unconfirmed, and three manuscripts were withdrawn the next day after a sign error.
- Checking takes years. A twelve-referee panel needed four years to reach “99% certain” on a single famous proof.
- The field is split between awe, grief, rejection and adaptation. Cédric Villani speaks of the most radical change in four centuries, and still tells his students the path matters more than the result.
- The paradox: we need more mathematicians to verify and digest AI results, at the moment AI makes the job less attractive to them.
- The expert who survives moves from producing answers to verifying, understanding and explaining them.
What happened in October 2026
Facts first. On 6 October, OpenAI published 722 preprints covering claimed progress on 372 problems in geometry, algebra, computer science and other fields (Retraction Watch). An internal model had been pointed at about 4,000 problems.
For Duminil-Copin, 16 of the questions he had worked on for years were in the batch. The open problems he listed in his lectures, his papers and his grant applications, i.e. his research agenda for the next decade, disappeared in one upload.
How mathematicians reacted
My reading of this event is far from isolated. The reactions of the field fall into four postures, and the most respected names sit in all of them.
Awe. Cédric Villani, Fields Medal 2010, said on France Inter on 8 October that this is “certainly the most radical change in mathematics in at least four centuries, maybe even in 1,000 years,” and that no university in the world produces high-level mathematics at this pace (franceinfo). The computer scientist Scott Aaronson called it “surely one of the biggest days in mathematical history.” Levent Alpöge, a mathematician at Anthropic, called it “the most significant moment in mathematical history” (Nature).
Grief. Duminil-Copin, of course. He also reported that his doctoral students and postdocs are in a state of total panic (SFEIR, summary of the October reactions). Villani himself had described OpenAI’s Navier-Stokes announcement in September as “a cataclysm unlike anything mathematics has ever known” (OfficeChai).
Rejection. Michael Harris (Columbia) calls the process illegitimate and says he will not read the papers. Ursula Martin (Oxford) described it as “tossing the community a messy first draft and expecting us to clean it up” (Nature). A newly created Association for Human Mathematics wrote that releasing over 700 files at once “is not a demonstration of scholarship, but a demonstration of power,” and Terence Tao reposted the statement on his blog (statement).
Adaptation. Timothy Gowers declined to sign the open letter of the Fields medallists: for him, a flood of results can be good if it expands knowledge, and his worry is the talent pipeline (IBTimes). Daniel Litt (Toronto) finds the release excellent for mathematics, and regrets that OpenAI does not employ enough specialists in enough fields. François Charles, who heads the mathematics department of the École normale supérieure, told Libération that “the scientific adventure will not stop with AI, it will accelerate,” and that “the scientist remains fully responsible for the content produced, pressing a button is not enough” (SFEIR).
What strikes me is the point on which almost all of them agree, even when they disagree on everything else. Villani said it most clearly in September: he told his own doctoral student not to use generative tools during the thesis, because “the path is even more important in science,” and because during training “you have to work hard to reshape your mind, your brain” (AFP via Techniques de l’Ingénieur). The awe and the grief come from the same place. Everyone sees that the machine now produces results. Nobody has a plan yet for the understanding.
Are the proofs correct?
Partly. And nobody knows yet which part.
Three manuscripts were withdrawn the day after the release: a sign error invalidated one argument, and two other papers built on the same construction fell with it. OpenAI revised 14 more manuscripts. A spokesperson said that roughly 50% of the results were released unconfirmed (Retraction Watch).
You could think formal verification solves the problem. Lean is software that checks every step of a proof against the rules of logic. Some of the manuscripts come with a Lean proof. But Johan Commelin, who leads the Mathlib Initiative, warns that translating a statement into Lean “needs human experts to read the Lean code,” and he suspects some formalizations “prove weaker statements than what’s claimed in the paper” (Quanta Magazine). The machine checks the proof. A human still has to check that the machine proved the right thing.
Other papers are simply hard to read. Mihalis Dafermos (Princeton) says some are too vague to verify. Daniel Litt (Toronto) suggests treating them as “research findings” or “raw data” until someone turns them into mathematics.
That last idea is the healthiest reading I have seen. A result that nobody understands is a lead. It becomes knowledge when a human can explain why it is true.
How long will it take to check?
Here is a reference point. In 1998, Thomas Hales announced a proof of the Kepler conjecture, a problem posed by Kepler in 1611 about stacking spheres. The Annals of Mathematics set up a panel of twelve referees. After four years, they said they were “99% certain” the proof was correct, and they could not certify all the computer calculations. Full certainty came from a formal proof completed in 2014, sixteen years after the announcement (Wikipedia).
That was one proof, written by a human who could answer questions. OpenAI released 722 manuscripts at once, written by a model that cannot defend its reasoning at a conference.
The mathematicians are already saturated. “We simply don’t have the resources to correct the record,” wrote one topologist (Quanta Magazine). The Erdős problems website, a reference list of open problems, froze new comments and proof claims because nearly all submissions were unexplained AI proofs, and moderation at that volume was no longer sustainable (Thomas Bloom, on Terence Tao’s blog).
So the honest answer is: years for the important results, and some will never be checked. Proofs became abundant. Understanding stayed scarce.
Why mathematics fell first
Why mathematics before, say, plumbing? For a large language model, mathematics is a language. A proof is a sequence of operations on symbols, which is exactly what these models do best.
They also have limits. I picture what a model understands as a globe: a shape inside which it is very strong. Outside the globe, it struggles a lot more. Many of the problems that fell look like searches for a counterexample or a clever construction, the kind of creative search that suits a language model well.
Now extend the idea. Any job made of operations on language sits inside the globe. Law is operations on language. Accounting is operations on numbers and on law. So is a large part of research, consulting and software. If that is your expertise, the same flood is coming: more answers than your profession can check.
Chess already lived this
Something else was associated with the human mind for centuries: chess. Deep Blue beat Kasparov in 1997, and twenty years later AlphaZero beat Stockfish, the best classical engine, in a 100-game match.
Chess is still alive. Top grandmasters prepare their openings and get their ideas by working with neural network engines, and Magnus Carlsen cited AlphaZero as a source of inspiration for his play in 2019 (ChessBase). Whole openings came back into fashion because the engines found new ideas in them.
Humans and machines now co-learn chess. Mathematics is entering that stage, and the job of the expert changes shape with it.
The paradox: who will check if experts lose the joy?
Here is the part that worries me most.
To absorb AI results, mathematics needs more people who can verify, formalize and explain. And AI is making the job less attractive to exactly those people. Duminil-Copin’s biggest worry is what this means for young people deciding whether to pursue a career in mathematics (OfficeChai). A doctoral student told Quanta that constant AI proofs “breaks our system.” Some early-career researchers now call for a boycott of AI for new theorems (Quanta Magazine).
I understand them. If you spent fifteen years training your mind to find a proof, and a machine now finds it in three hours of compute, the pleasure of the craft takes a hit. Craig Kaplan (Waterloo) put the stake well: “The path they took was often more important than the final answer.”
In September 2026, 28 Fields medallists signed a declaration on mathematics and AI. Their core point: “solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight.” They are explicitly open to AI, which “offers the potential of enhancing and accelerating genuine mathematical study and understanding,” and they ask that the people in control of the technology steer it there (A Severe Misalignment of AI in Mathematics).
I agree with them. When you solve a problem, the path to the solution is often more important than the solution. A mathematician finds an elegant proof, we follow the reasoning, the reasoning prints itself in our brain, and then it helps us see other problems creatively. If we delegate everything to AI, we stagnate, because nobody can learn from a reasoning that nobody has followed.
We need more mathematicians, working with AI
So my conclusion goes against the mood of the moment: the world needs more mathematicians, and they need to work with AI.
The new jobs are already visible:
- Formalizers. People who translate proofs into Lean and check that the formal statement matches the claim. Mathlib, the community library of formalized mathematics, held over 210,000 theorems in May 2025 (Wikipedia). It needs far more hands.
- Expositors. People who turn raw AI output into mathematics others can learn from. The Erdős problems site now prioritizes clear human write-ups over a scoreboard of solved problems.
- Problem choosers. Terence Tao argues that proofs become abundant and understanding becomes the scarce resource, so mathematicians should focus on choosing the right problems, verifying results and digesting them (Phemex News, summarizing Tao).
- AI co-authors. DARPA’s expMath program funds an “AI co-author capable of proposing and proving useful abstractions,” because breaking problems into lemmas and proving them is “slow, effortful, and iterative” (DARPA).
Some mathematicians already live this. Ahmed Bou-Rabee (Penn): “AI has allowed me to do things I was never able to do before” (Quanta Magazine).
My prediction: a lot of mathematicians will take the path I took a few years ago, toward communication. Someone in my community asked me why the AI could not explain its own work. It will explain a lot. But an explanation creates trust when someone takes responsibility for it: a name, a face, a reputation that can be lost. That part stays human.
I left research for marketing in 2022, and I explained why in Why I Quit AI Research to Work in Marketing. The skill transferred well: understand a hard thing, then make other people understand it. Since then, the content strategies I ran for clients generated more than 400 million organic views in two years (last count January 2026).
“Where was the empathy when machines replaced workers?”
When I shared the mathematician’s story with my community, the most common objection was sharp and fair: where was the empathy of these people when machines replaced factory workers? Several people told me their whole department was replaced by software ten years ago, or that a robot took their factory job thirty years ago.
They are right about history. Office workers cried very little for factory workers, and knowledge workers now get the same treatment.
I keep asking for empathy anyway, for a practical reason. If you cannot understand why someone suffers when AI takes the thing they spent twenty years building, you cannot read what AI is doing to humans. And if we mock the experts who suffer, we lose the very people we need to check the machines. The worker replaced in 1990 deserved the same understanding.
The real risk: a dark age
I am very optimistic about the future. I also think there is a real reason to be afraid of AI, and it is far from the robots of science fiction.
The risk is a dark age where we happily delegate our thinking. If abstract problems are solved by AI alone, mathematicians become AI technicians who check that the machine solved the problem, without understanding how.
Some humans will still understand. They will probably sit inside the companies that own the models. That is already the trend: in 2023, industry produced 51 notable machine learning models, against 15 for academia (Stanford AI Index 2024). In Europe, a public lab is happy with a grant of a few hundred thousand euros, while the AI labs of the big tech companies spend hundreds of millions.
So refuse the obscurantism. Being good, being smart and working on abstract concepts stays necessary for humanity, whatever the machines do.
What to do if you are the expert
I am not a guru, so here is just the reading grid I used for myself.
- Be gentle with yourself. The pain is normal. It shows that your work was part of your identity, which is a noble thing.
- Take care of your material needs first. I have three children, so money is part of my equation too.
- Become the verifier. Learn the tools that check machine output in your field (Lean for mathematicians, tests and audits elsewhere).
- Become the explainer. When answers are abundant, the person who makes them understandable is scarce.
- Work in public. When your result can be generated, your understanding and your judgment become what people trust.
Someone in my community quoted Yann LeCun: the invention of the boat reduced the importance of swimming, and it allowed the discovery of new lands. I could not find the original source of this quote, so take it as a metaphor. It is a good one.
So, your turn. Which part of your expertise would you bet your career on, and who in your field will check the machines?
The result now belongs to whoever owns the compute. The understanding belongs to whoever does the work of checking and explaining it.
FAQ
Are OpenAI’s AI-generated math proofs correct?
Partly. Three of the 722 manuscripts were withdrawn the day after release because of a sign error, 14 were revised, and OpenAI said about half of the results were released unconfirmed. Lean verification helps, but a human still has to check that the formal statement matches the claim.
How long does it take to verify a major mathematical proof?
Often years. The Kepler conjecture proof took a twelve-referee panel four years to reach “99% certain,” and a full formal proof was completed sixteen years after the announcement.
How did mathematicians like Cédric Villani react to OpenAI’s math papers?
With a mix of awe, grief, rejection and adaptation. Villani called it the most radical change in mathematics in at least four centuries, Duminil-Copin said his field was annihilated, Michael Harris refuses to read the papers, and Timothy Gowers sees value if the results expand knowledge.
Will AI replace mathematicians?
AI produces proofs at scale, and the field now needs more people to verify, formalize, explain and choose problems. The job changes shape, as it did in chess.
How can an expert stay relevant when AI produces the answers?
Move toward verification and explanation: learn the tools that check machine output in your field, explain results others cannot read, and build a public reputation for your judgment.
Annex: the original video
I first shared these thoughts, raw and unedited, with my community on TikTok:
Dr Louis Korczowski has a PhD in AI and neuroscience and spent ten years in brain-computer interface research before moving to marketing. He now helps SaaS and AI founders win the idea market before the product market at STATUR.