AI and mathematics: when checking costs more than producing

AI and mathematics: when checking costs more than producing

OpenAI released 722 math manuscripts at once and mathematicians can't keep up. Their crisis shows where the work goes: choosing, checking and understanding.

Concepts
10 min read

On October 6, 2026, OpenAI pushed 722 manuscripts of mathematical results produced by one of its internal models to GitHub. The next day, a sign error in one of them forced three to be withdrawn. Two days after the release, Hugo Duminil-Copin, 2022 Fields medallist, described a "monumental shock" on Swiss public broadcaster RTS: sixteen of the manuscripts touched on his own research field. In a few weeks, the relationship between AI and mathematics changed in kind.

Producing a proof has not become free. For Navier-Stokes, OpenAI ran about 10,000 agents in parallel, at a cost of millions of dollars. But per result, it is fast and far cheaper than a researcher's work: the OpenAI release reports an average of three hours of compute per result. What is becoming scarce is researchers' time: to check those proofs, to understand them, and upstream, to choose the problems worth asking. The day after the release, only about 42% of the top-line results had been checked by a computer.

Mathematicians are not the first to go through this. Developers did before them, and expert professions, from law to audit, are next. Everywhere, value is moving from production to what surrounds it: choosing what to produce, checking, understanding, judging what applies. Mathematics is the purest case, because checking can be automated there and the rest cannot.

What did OpenAI and Anthropic actually announce in mathematics?

Four announcements came within three months. They do not all say the same thing, and several were summarized badly in the press.

DateAnnouncementWhat was actually done
August 10, 2026Anthropic, Riemann zeta functionA bound on the share of zeta zeros raised from 41.6% to 67.2%. Not a proof of the Riemann hypothesis.
September 4, 2026Anthropic, Fermat's Last TheoremThe existing human proof translated into Lean and checked by computer. Not a new proof.
September 8, 2026OpenAI, Navier-Stokes equationsA finite-time blow-up result on a Millennium Prize problem, obtained with about 10,000 agents.
October 6, 2026OpenAI, GitHub release722 manuscripts, 372 result families, about 4,000 problems given to the model.

Riemann: a better bound, not the hypothesis

In August, Anthropic announced that Claude had improved a longstanding lower bound: the share of zeros of the zeta function on the critical line goes from 41.6% to 67.2%. It is a real result, formalized in Lean. It came out of a failed attempt to prove the hypothesis itself, and Anthropic warns: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis."

Fermat: a verification, not a proof

In September, Anthropic published the first complete computer-checked proof of Fermat's Last Theorem. Claude did not discover a proof: it rewrote Andrew Wiles's proof in the Lean language, following a simplified version by Darmon, Diamond and Taylor. The work is still huge: 13 million lines of Lean in 11 days, more than five times the size of the field's reference library, according to Anthropic. Kevin Buzzard, who leads the human Fermat formalization project at Imperial College, praises "this extraordinary autoformalization achievement", while carefully adding "which Anthropic researchers say only took 11 days". His project has £1 million over five years, since 2024.

Those 11 days were not free either: Claude produced about six billion output tokens, the units of text a model bills for. As the model is internal, the real cost is not published; at the public price of a comparable model, TNW puts it at $300,000, as an illustration. Humans did not disappear, though. According to Anthropic, human mathematical input was limited to "occasional high-level instructions" from one researcher, who for instance pointed to which part of the proof to prioritize. Few messages, but they set the direction.

Navier-Stokes: a result and a controversy

On September 8, OpenAI announced a result on one of the seven Millennium Prize problems. The day before, mathematician Tristan Buckmaster had published related results and raised a question in public: he and his co-author had put all their drafts into Codex, OpenAI's coding tool. Had the model accessed them? He was told no, without an answer to his question about training. Buckmaster adds: "I am not accusing anyone of anything."

According to MIT Technology Review, OpenAI's Sébastien Bubeck and Mark Chen said the team only succeeded by running about 10,000 agents concurrently, at a cost of millions of dollars. Buckmaster and his co-author, Levent Alpöge, had worked on the problem for almost a year with public models, without a full solution. If OpenAI's models were trained on their work or had access to it, journalist Grace Huckins notes, that would suggest the two researchers' work was essential to the result. The same article drew a wider conclusion: "Mathematics is quickly becoming the province of frontier AI companies".

The October 6 release: raw data

OpenAI presents its 722 manuscripts as "a form of raw research data, rather than polished academic papers" and admits that "some of the unformalized results could have issues". AGMAI, an independent advisory group of mathematicians set up after OpenAI approached some of its members, summed it up in one sentence: "This release is the beginning, not the completion, of the process of human understanding (…)". The work of understanding starts after publication, and it falls to humans.

Lean, the proof assistant that checks AI proofs

A proof assistant is software in which you write a proof that the computer checks step by step. If an argument is flawed, it rejects the proof, just as a compiler rejects a broken program. Lean, one of the most widely used today, is an open-source language with a community library, Mathlib, that holds over a million lines of formalized mathematics.

For a developer, the analogy is obvious: Lean does for a proof what the compiler and the tests do for code. That is what lets AI companies release hundreds of results while claiming that at least part of them is guaranteed. For the OpenAI release, the share was computed on October 7: 300 top-line results out of 719 formalized, about 42%. The rest was not formalized in Lean at the time of publication.

Lean has a limit, though: it guarantees that a proof is correct, not that it is understood. The Jacobian conjecture, disproved in July 2026 with Claude Fable, an Anthropic model, is a good example. Terence Tao, one of the world's best-known mathematicians, notes that the counterexample can be checked "by a brief calculation". He adds: "I do not have a completely satisfactory geometric explanation for this miracle." The result was correct and checked. Understanding it was a separate piece of work, which other mathematicians took on afterwards.

Why mathematicians are reacting so strongly

Because their discipline rests on a simple idea: a result is only worth something if you understand why it is true. Matthieu Alfaro, a professor at the University of Rouen Normandy, puts it this way (our translation): "The day you prove a conjecture is true, it's no surprise. What is interesting is how you proved it."

Hugo Duminil-Copin spoke to Swiss broadcaster RTS on October 8, two days after the release: "I even cried, to tell you everything", "we are the day after". The same day, Cédric Villani, Fields medallist and former French MP, spoke on France Inter radio of the "most radical change in mathematics in at least four centuries", adding that "no university in the world produces high-level mathematics at such a pace".

Collective reactions came before the October release. On September 11, 25 Fields medallists published a statement, "A Severe Misalignment of AI in Mathematics", since signed by 28 of them, saying that "the goals of the AI companies and the goals of the mathematical community are severely misaligned". It has since been backed by more than 8,000 mathematicians, according to the French think tank Telos. The board of the French Mathematical Society saw the Navier-Stokes result as "probably a major advance", while denouncing "bibliographic plundering" and a "total absence of collective strategy". After the release, the Association for Human Mathematics, which has over 1,000 members, went further: "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power". It urges mathematicians to stop working with OpenAI.

There is also a very concrete volume problem. According to a preprint, mathematics submissions to arXiv, the main preprint platform, rose 33.5% between January and August 2026 compared with 2025. Matthieu Alfaro observes that journals receive "twice as many papers" as a few months ago, reviewed by researchers who were already struggling to keep up.

Then there is the question of means. "What is clear is that very few mathematicians will have resources of that scale", Javier Gómez-Serrano told MIT Technology Review, about the means OpenAI put into Navier-Stokes. Christophe Besse, director of CNRS Mathematics in France, sees a matter of "scientific independence" and wants to keep "a public and academic capacity to develop, study and use these tools" (our translation). It is the same dependency question that sovereign AI raises for companies.

Are they defending their science or their jobs?

The question comes up, and the facts support two readings.

First reading: they are defending science

Mass-produced proofs are often unreadable, even for specialists. Three were withdrawn the day after the release, and more than half were not formalized in Lean. The Fields medallists' statement recalls that "the most precious resources of our profession are students and ideas", and that mass production of true-or-false statements "could destroy fertile ground". Duminil-Copin speaks of "bulldozing" problems: when a question is settled without anyone understanding how, the ideas that would have emerged along the way never appear.

Second reading: they are also defending a profession

Matthieu Alfaro says it plainly: research projects swept away by AI, like Duminil-Copin's, fund PhD and postdoc positions. And competition changes scale (our translation): "If I tell my student not to use AI while ten others use it all day, by the time he has written three papers, they will have written thirty."

Our reading

We think the question is partly the wrong one. In mathematics, the science and the profession are the same thing: understanding. The useful test is what each person does with a result that is correct but incomprehensible: reject it, or make it the raw material of new work. Duminil-Copin himself does not reject the tool: "Artificial intelligence is revolutionizing the way we discover ideas", he says. He argues for a use directed by a community that would choose the problems.

That work is not limited to checking. MIT Technology Review notes that experts have long seen "research taste", the ability to choose promising research questions and directions, as a major obstacle for AI. Terence Tao points out there that in pure mathematics, problems are usually posed not because anyone desperately wants the solution, but because the effort to solve them moves the whole field forward. Choosing the problems, setting the direction, judging what is useful: that is where researchers' value is moving, as much as to verification.

Other French voices are clearly more open. François Charles, head of the mathematics department at the École normale supérieure, says he is "not at all pessimistic" about the science: "As scientists, we have to be more ambitious." Stéphane Mallat, CNRS gold medallist, says in the same AFP report that "we will need mathematicians because mathematics remains central to our ability to understand things clearly". And Terence Tao offers the most useful rule in the whole debate, in a summary of his views that he reviewed: only rely on AI where you are able to red-team its output.

The same bottleneck, from code to law

Developers got there first

In July 2026, Claude was writing about 80% of the code merged at Anthropic. In its main repository, merged code changes went from 2,500 to more than 30,000 a day between May and September, to the point that Anthropic moved that repository off git to an in-house tool, Scion. The team that built it now checks the intent of each change rather than every line. The team puts it this way: "the role of the engineers was to tell Claude "no.""

Across GitHub, 1.3 billion pull requests were merged in the twelve months to the end of September 2026, up from 538 million the year before, and more than 40% are now started by an agent. Someone has to review them. According to telemetry from Faros AI, a measurement tool vendor, median review time grew fivefold, with 28% more bugs per pull request. In Stack Overflow's 2026 survey, 48% of developers say they trust AI when they can easily validate its answers: trust follows verification.

Security has been going through the same thing since the spring. By November 2, Anthropic had reported 5,087 vulnerabilities found with its models to 680 open-source projects, and only 629 were fixed. We described in our article on Claude Mythos how the bottleneck became fixing, not finding.

Expert professions follow, without Lean

Code has tests and mathematics has Lean. Law, audit and consulting have nothing equivalent: verification there remains human, slow and expensive. When it is skipped, the damage shows.

Damien Charlotin, a French legal scholar and researcher, tracks court decisions involving content invented by AI, such as case law that does not exist. As of November 12, 2026, his database lists 2,292, including 912 involving lawyers and 14 in France. On June 4, the administrative court of Grenoble imposed a 200-euro fine for abusive proceedings on a claimant whose brief of over 300 pages was "manifestly generated with the help of artificial intelligence" (our translation), citing court decisions that do not exist.

In consulting, in 2025, Deloitte had to refund 97,587 Australian dollars to the Australian government, on a contract of about 440,000, for a report full of non-existent references. In 2026, GPTZero, a company that sells an AI text detector, checked a KPMG report: 5 accurate citations out of 45.

Inside companies, the effect is more diffuse but measurable. According to a Workday survey of 3,200 employees, "nearly 40% of AI time savings are lost to rework": correcting, rewriting and checking what it produces. The advertised gain of an AI tool is not its real gain.

What a company can take from this

Mathematicians are living a few months ahead of what many legal, finance and technical departments will go through. Their crisis offers five practical lessons.

  1. Put experts upstream and downstream. Choosing the right questions, framing the request, judging whether a result applies to the real case: that is where their time is worth most, like the few instructions from Anthropic's researcher on Fermat. Producing is no longer the core of their value.
  2. Count verification time in the return on investment. A tool that produces a report in ten minutes and needs two hours of checking does not save the time it advertises. The honest measure is the net gain, as the Workday survey shows.
  3. Prefer outputs that can be checked. An answer that cites its sources, a calculation done by a program rather than a model, code that comes with tests: anything that makes verification mechanical lowers its cost. That is why we advise using as little AI as possible when a classic algorithm will do.
  4. Review according to risk. Reviewing every line no longer holds at these volumes, and reviewing nothing does not either. Vercel merges 58% of pull requests in its largest repository without human review: an AI classifies each change against risk criteria set in advance, and sensitive cases still go to a human reviewer. A contract, a board memo or a client quote can follow the same logic.
  5. Keep someone who understands. Validating that a result is correct is not always enough: it has to be explained to a client, a judge or an auditor. Junior professionals are the most exposed, like the PhD students Alfaro talks about, and that is what AI training should work on first.

Frequently asked questions about AI and mathematics

Did AI prove Fermat's Last Theorem?

No. In September 2026, Anthropic published the first complete computer-checked proof of Fermat's Last Theorem. Claude translated into the Lean language, in 11 days according to the company, Andrew Wiles's human proof in a version simplified by Darmon, Diamond and Taylor. It is a formalization, meaning the automatic verification of a proof that already existed, not a new proof.

What is Lean in mathematics?

Lean is an open-source proof assistant: software in which you write a proof that the computer checks step by step, the way a compiler checks a program. Its community library, Mathlib, holds over a million lines of formalized mathematics. AI companies use it to guarantee that their models' proofs are correct, without making them understandable.

How many mathematical results did OpenAI release?

On October 6, 2026, OpenAI pushed 722 manuscripts grouped into 372 result families to GitHub, from about 4,000 problems given to an internal model. Three manuscripts were withdrawn the next day after a sign error was found, bringing the total to 719, then two were added on October 26 (721). On October 7, about 42% of the top-line results were checked in Lean. The often-quoted figure of 377 is the number of the last family, not a count.

Will AI replace mathematicians?

Not according to the mathematicians who have spoken out, including the most worried. Stéphane Mallat says mathematicians will be needed to understand, and François Charles thinks some roadblocks may be overcome. What changes is the job: less time proving, more time choosing problems and research directions, checking, sorting and understanding results produced by machines. Matthieu Alfaro already describes a multi-speed research world, between those who use AI all day and those who refuse it.

Why are mathematicians criticizing OpenAI?

They criticize OpenAI for releasing results in bulk, faster than the community can check and understand them. The Association for Human Mathematics calls it "a demonstration of power" rather than scholarship. Mathematician Tristan Buckmaster also publicly asked whether his team's drafts, stored in OpenAI's Codex tool, had been used by the model. He was told the model had not looked up user data, with no answer about training.


Mathematics has an advantage most professions lack: a tool, Lean, that says without ambiguity whether a proof is correct. Even with it, the community is overwhelmed, because checking is not understanding. In a firm, a finance department or a legal team, there is no Lean. Before rolling out a tool that produces faster, two questions come before everything else: who will decide what it should produce, and who will check what it produces, and how long will that take?

Partager :

Related Articles

Pick a number between 1 and 10: why AI always says 7Concepts

Pick a number between 1 and 10: why AI always says 7

We asked seven AI models for a number between 1 and 10, 14,000 times. Five answer 7 almost every time, whatever the temperature. Where does this bias come from?

9 min
AI Distillation: Everyone Blames China, Everyone Does ItConcepts

AI Distillation: Everyone Blames China, Everyone Does It

OpenAI and Anthropic accuse Chinese labs of distilling their models. Yet Meta, Google, Mistral and even Musk's xAI distill too. Here is what we actually know.

11 min
GPT-6.2 vs Claude Fable 5.2: Same Price, Not the Same CostAI

GPT-6.2 vs Claude Fable 5.2: Same Price, Not the Same Cost

GPT-6.2 Astra and Claude Fable 5.2 both cost $8 and $40 per million tokens. At equal scores, a Fable task still costs 3.6 times more. Independent data inside.

9 min

Interested in our solutions?

Discover how Ask This Guy can help you accelerate with AI

Book a Demo