Notes & Working Papers
2026.11.18
v1.0

Machines Prove, Humans Accept

AuthorROME THORSTENSON
FiledNOVEMBER 2026
Reading14 min
StatusPUBLISHED

On 8 September 2026, OpenAI announced that one of its systems had solved a Millennium Prize problem. The proof showed that, under a suitable smooth external force, solutions to the Navier–Stokes equations can blow up in finite time, and it came with a formalization in Lean. The Clay Institute, which awards the prize, wrote of "the announcement that the Navier-Stokes problem has apparently been settled" and said its own process is "deliberately unhurried." No one has disputed that Lean's kernel accepted the proof. Mathematicians have disputed whether the theorem OpenAI proved is the problem they care about.1 Bastounis, Circelli and Hansen argue that the Lean proof "does not correspond to" the argument in OpenAI's paper.

Days later, 28 Fields Medalists signed a letter against the AI companies' push into mathematics. My view has three parts.

  1. Machine proof is useful for mathematics.
  2. Whether a proof is correct and whether a result is accepted are different questions; only acceptance requires human understanding.
  3. The mathematicians pushing back are right that mathematics—or rather, human mathematics—still needs people. What it needs them for is understanding, which is a prerequisite for accepting a result.

None of this makes the change easy. Proving a hard theorem has long been the strongest signal mathematics had: evidence that someone had new insight, new methods, and the ability to produce more. Hiring, prizes and reputations rest on that signal. A machine proof undermines this process. Mathematics will undoubtedly spend years working out what to measure instead.

This post is inspired by a paper on rigor and acceptance in mathematics that I am currently workshopping. If you would like to read a draft or have comments, reach me directly or here.

§ 01 Recent

Recent events

Fig. 1 follows AI work on Erdős problems in two catalogs. The upper panel preserves Tao's contribution records through June. The lower panel uses a newer catalog, Plasma, to follow numbered problems with a first full solution credited with substantive AI input. The catalogs have different units and inclusion rules, so their totals stay separate.

0 40 80 Oct ’25 Jan ’26 Apr Jul Oct 19 Nov 0 40 80 Oct ’25 Jan ’26 Apr Jul Oct 19 Nov a Tao archive · cumulative full-solution entries 19 standalone 15 alongside literature 17 building on literature 37 with humans 88 archive ends 30 Jun one problem can recur b Plasma catalog · first full solutions with substantive AI input cumulative numbered problems · since 1 Oct 2025 · catalog acceptance 53 98 16 Nov · #1017 scope caveat¹ Two catalogs, two units. No counts are joined or added across panels. ¹ The newest claim answers the catalog’s estimate question, not every formal benchmark target.
Fig. 1 Two retrospective catalogs, with different counting rules. (a) Tao’s archived full-solution contribution table: 88 dated system–problem entries by 30 Jun 2026. A numbered problem can occur more than once; date ranges use their end dates, and undated records are omitted. (b) An audit of Plasma’s current accepted full records: 98 distinct numbered problems whose first accepted full record in the catalog falls after 1 Oct 2025 and credits a model with mathematical work. Later formalizations, earlier known solutions and unclear attributions are excluded. The line uses catalog publication dates with documented corrections, and ends at its latest qualifying record on 16 Nov. “Accepted” is the catalog’s classification. Neither panel establishes mathematical consensus or an exhaustive world total. Snapshot checked 19 Nov 2026. Sources: Tao and contributors; Plasma AI and Erdős record contributors (catalog writing CC BY 4.0). Audit decisions, date corrections and sources accompany the figure.

Tao's dated full-solution entries stood at 40 at the end of March and 88 at the end of June. The largest group was AI working with humans: 37 of the 88 entries. The newer audit counts 98 numbered problems in its reporting window, 53 through June and 45 afterward, with the latest record dated November 16.2 Notably, Anthropic's Fable was publicly released in June, and OpenAI's Astra was released after that.

Formalization has sped up as fast. Fig. 2 shows Mathlib's theorem declarations and repository lines, followed by seven formalization projects. The newer cases include Fermat's Last Theorem, whose integrated Lean proof built on August 19, and integer superharmonic matrices, whose proof targets closed on November 6. Many of the foundations laid down in earlier proofs are reused in later ones. The sphere-packing durations cover the final machine runs, not the preceding human work; these projects are not a controlled comparison of human and machine speed. AI can write formalizations fast.

a Mathlib growth · latest available observations, 19 Nov 2026 Theorem declarations 100k 200k 300k 295,033 Repository lines 1m 2m 2020 2021 2022 2023 2024 2025 2026 2.506m all repository files dashed gray: Mathlib 3 · blue: Mathlib 4 · theorems and lines have separate scales b Selected formalizations elapsed / final run project source lines 2004 2010 2016 2022 2026 Flyspeck (Kepler) 11 years ~500k Liquid Tensor 19 months ~85k PFR 22 days 11,094 Sphere packing, dim. 8 5-day final run ~80k Sphere packing, dim. 24 2-week final run ~200k¹ Fermat’s Last Theorem 12 days² ~13m³ Superharmonic matrices 51 days⁴ ~234k Open markers: completed-by bounds. Flyspeck’s start is known only to the year. ¹ Combined 8D + 24D. ² Through full build. ³ Includes generated code. ⁴ Through proof closure.
Fig. 2 (a) Documented theorem declarations and repository lines on separate scales. Gray is Mathlib 3; blue is Mathlib 4, whose port initially overlapped the old library. The theorem count was observed on 19 Nov 2026; the latest Git line count covers commits through 18 Nov 2026, 23:47 UTC. Repository lines include non-Lean files, blanks and comments. (b) Seven formalizations, with human-led projects in gray and machine final runs in blue. Bars show documented start-to-completion intervals; open markers give only the latest known completion bound. Sphere packing rests on earlier human work. Code sizes use different snapshots and counting rules: Flyspeck uses HOL Light and Isabelle; the rest use Lean. PFR is our count of all tracked Lean files at its 5 Dec 2023 snapshot. Fermat includes generated boilerplate; the reported count is about 10.5 million lines without it. Sources: Mathlib statistics, pinned Git statistics, Flyspeck, Liquid Tensor, PFR snapshot, sphere-packing maintainers, combined sphere-packing size, Fermat timeline, superharmonic-matrices manifest and reported run sizes. Data and counting definitions.
§ 02 Correct

Correct and accepted

Mathematicians judge whether a proof is correct by checking it. A referee reads an informal proof and judges whether its steps hold. For a formal proof, a kernel checks every step, at least as thoroughly as any human referee.3

A result is accepted when a community treats it as established: its members cite it without hedging, build on it and teach it. Acceptance is a decision people make on evidence. It requires that they understand what the result says, though not necessarily the proof.

Correctness and acceptance do not always go hand in hand. Consider the following cases:

In 2016 Heule, Kullmann and Marek showed that the numbers 1 to 7,825 cannot be colored red and blue without leaving some Pythagorean triple all one color. The proof is a computer-generated certificate of nearly 200 terabytes that no human will ever read.4 Mathematicians accept it because they are satisfied of three things: the formal statement says what they mean, the checker checks the logic they take it to check, and the checker can verify the whole proof successfully.

The unit-distance disproof was accepted before anyone formalized it. Mathematicians outside OpenAI checked the informal proof by reading it. The Lean proofs came weeks after the verdict was in.

The Navier–Stokes proof has passed a kernel and is not yet accepted as a solution of the problem. Whether its theorem is the one mathematicians meant—which is to say, whether the problem has been solved—is a question only people who understand the problem can settle.

Acceptance can also stall with no machine involved. Shinichi Mochizuki's claimed proof of the abc conjecture appeared in 2021. Most experts outside his circle have not accepted it, since Peter Scholze and Jakob Stix argued in 2018 that a key step fails. A project working toward a Lean formalization, publicly launched in 2026, is now examining that step. Acceptance waits on someone outside his circle coming to understand it.

A formalization that succeeds need not bestow understanding. Starting in December 2020, Johan Commelin led a team formalizing one of Scholze's hardest theorems in Lean, at Scholze's invitation, to check a step he doubted. Half a year in, asked who besides the authors now understood the proof, Scholze answered, "I guess the computer does, as does Johan Commelin." He also wrote that he could not read the Lean proofs himself.

Read by people Machine- checked Statement agreed Accepted Liquid Tensor Experiment, 2022 only Scholze Lean Unit-distance disproof, 2026 outside after Boolean Pythagorean triples, 2016 200 TB checkers abc conjecture, Mochizuki disputed not yet few Navier–Stokes blowup, 2026 under way Lean disputed not yet yes no partial, disputed or under way
Fig. 3 What each case had. Blue marks the statement and acceptance columns.

Machines now prove theorems and explain proofs, but they cannot accept a result on our behalf. Accepting a result means citing it, building on it and teaching it, and no one can do that with a statement they do not understand. A machine's explanation has done its work only when a person understands the result well enough to use it, teach it or find it wrong.

§ 03 Why

Why are mathematicians pushing back?

The Fields Medalists' letter makes two charges.

The first is that "the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics." Famous problems, the letter says, have served "as landmarks and lighthouses," and solving one has been "a certain sign of new insights and interesting methods" that the community then studies and teaches. By the letter's account, a machine that settles a famous problem without new methods removes a landmark and gives the community nothing to study. A student assigned a problem "with the core intention of developing skills" gains nothing from a machine's answer to it. And the letter warns that mass-producing "true/false" statements ever faster "could destroy fertile ground."

I think this first charge is tragically flawed. A machine proof does not remove a landmark. The problem is still there, with a solution attached that can be studied—or deliberately ignored. On the day the unit-distance disproof appeared, Will Sawin derived an explicit bound from it, and within a month a construction of the same kind disproved a second Erdős conjecture, on sum-product sets over the real numbers. Students are still assigned problems whose answers are known; every textbook exercise is one. Machine-produced "true/false" statements are fertile ground for anyone who studies them. The charge treats solving as the source of mathematics' value; the letter's own premise, that solving is a proxy for understanding, refutes it.

The second charge is that "the goals of the AI companies and the goals of the mathematical community are severely misaligned." The companies want solved problems as evidence of what their models can do. The mathematical community wants what solved problems once sowed: new methods, understood by people and passed on. Without mathematicians willing to integrate machine results into the canon, the letter says, "the crucial human transmission chain between mathematicians would be lost."

As a sidenote, that everyone and their brother is jumping on the bandwagon of building yet ever-complexifying harnesses to demonstrate they can harness this intelligence is a lamentable byproduct. It's unclear, aside from demonstrating they can build harnesses well, what exactly most of them hope to achieve, but I digress.

This second charge seems to me reasonable, unfortunate and entirely possible. Given the flaw in the first charge, though, be careful not to overestimate how misaligned AI labs and mathematicians are. The methods mathematics reaps from its problems exist to solve problems, and we can now solve them faster and more reliably than ever before.

Yehuda Rav imagined a machine, PYTHIAGORA, that answers any well-posed mathematical question with "true" or "false."5 (Remember, we do get informal and formal writeups out of these systems, however impenetrable they might be.) If proofs existed only to establish truth, the machine would make proving redundant. Rav's answer was "no boon but doom": such a machine "would have dealt a death blow to mathematics, for we would cease having ideas and candidates for conjectures," because proofs "are the heart of mathematics, the royal road to creating analytic tools." William Thurston put the purpose of mathematics more directly: "The measure of our success is whether what we do enables people to understand and think more clearly and effectively about mathematics."

Talk about the purpose of mathematics can hide how mathematics measures success in practice. Jobs, tenure, prizes and opportunities go to whoever proves the most interesting and important results. That makes credit assignment the largest gap in how AI labs work with mathematics. A proof found by a model, run by a company and built on published work the model does not cite, can obscure who contributed what. Formal proof could help: if the results a Lean proof uses carry attribution, we can trace those contributions even when a machine assembles the final proof.

The letter also says what problems are for: "solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight." Scholze's endorsement of the Leiden Declaration puts it more briefly: "The goal of mathematical research is human understanding of mathematics." (Yes, pretty much everyone who writes about the purpose of mathematics defines mathematics circularly. That's just how it is.6)

But it should be said, explicitly: AI labs have no intention or goal of undermining human understanding of mathematics. They have a different goal: evaluating systems that have useful applications. Many mathematical problems are economically useful.7 The unfortunate conundrum is that contemporary mathematical practice and education are organized so that easily knowing the answers threatens both. This is, ultimately, a critique of current mathematical practice, in particular that it is already failing to achieve its goal directly. It gets there (human understanding) mainly as a happy byproduct of making the business about proving things.8 Many people, mathematicians included, would argue—as the letter's first charge implicitly does—that human understanding is for proving things, not the other way around. In some sense, this is a good-hearted, Goodharted defense.

If solving problems is the end goal of math, then machines proving theorems left and right displaces humans. But the point is that mathematics is not about solving problems. It's about understanding—where machines can only help. The harm comes from laziness and incentives. People and companies both follow what institutions reward, whether that is a final exam and tenure or evidence and outcomes of better AI. Benchmarks make the proxy the target, and attention and money follow the target. Unfortunately, results announced "in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others" as the letter puts it are results no one can yet learn from.

Much of the Fields Medalists' letter is about students. "The most precious resources of our profession are students and ideas," it says. The letter has two audiences. One is the current generation of mathematicians, many of whom are alarmed. Hugo Duminil-Copin, one of the signatories, wrote in August that AI-generated discoveries "nuke the mathematical landscape, making it increasingly difficult to inhabit after each blast," and that "the social consequences for our community, especially for the younger generation, are devastating." He asked, "What are we supposed to tell young PhD students entering a field?" The other audience is that younger generation. A student weighing years of training is watching AI systems solve extraordinary problems at an ever increasing rate, and may well doubt that proving is still worth learning.

I read the signatories as forming a unified front not to defend their own conceptions of what math ought to be but because they care deeply about the field and the people entering it. That is an admirable goal. Telling students that machine proof harms mathematics will not keep them in the field, because they're watching machines solve important problems ever faster. Instead, they attack the root of the problem by articulating how the temptation to bypass the fruits of math, the understanding, undermines the work of generations of mathematicians. The letter's premise that solving problems is a proxy for understanding supports that answer, and I would lead with it. A result counts only once someone understands it well enough to accept it—whatever machines can do.

It is my conviction that one can prove things without humans, but that for a mathematical community to accept a result, humans must understand it.9 In other words, math doesn't need people, but people benefit from math.

§ 04 Replies

Replies to objections

Careers rest on proving theorems. The cost is real. Hiring, tenure and grants count theorems, and people whose theorems a machine could have proved faster will compete for fewer jobs. Institutions can start counting understanding, teaching and exposition, but only if mathematicians say plainly that understanding is the point. I expect incentives to follow, slowly. The transition will be painful.

Machines will understand too. Perhaps. They already explain proofs and write informal versions of formal ones, which helps people understand. As long as the community that cites and teaches a result is made of people, those people have to understand it themselves.

Lean can certify the wrong statement. It can, and Navier–Stokes may be a case. Checking a formal statement against the intended one is work for people who understand the mathematics, which is a reason to train more of them.

Machine proofs can be unreadable, not just unsurveyable. Klowden and Tao note that AlphaProof's Olympiad solutions "contained numerous redundant or inexplicable steps but nevertheless were formally verified to be correct solutions." I assume that enough effort, from people or machines, can make most such proofs understandable. I have not seen that assumption tested. It is an open question when and how to choose to interpret machine proofs or attack the problem unaided.

Machines will choose the problems. I know of no conjecture proposed by an AI system that mathematicians regard as profound. That may change soon. If it does, judging which problems deserve work becomes a larger share of what mathematicians do.

Attribution. Credit is a separate, critical problem. The Navier–Stokes announcement, for example, set off a dispute over credit with Tristan Buckmaster and Levent Alpöge, who had reached a result on the forced Euler equations weeks earlier; OpenAI acknowledges their priority on that result. It's an open problem.

§ 05 Conclusion

Conclusion

The essential question is how much proving practically matters to understanding. The preceding millennia have shown that proving is a great way to develop understanding, and to demonstrate it. If people who hand their proving to machines stop understanding the results, they will have let themselves down by depending on the machines, and machine proof will have carried a real cost. In practice the question is whether the temptation to hand off proofs will undermine a generation of students now entering mathematics. I hope not; I take up a version of this worry in Everyone Who Can Appreciate a Symphony.

I am also thrilled to study a future in which AI helps us explore the primordial mathematical reality of our universe.10 Who knows what fascinating conjectures and problems we will find together?

Notes

  1. N · 01

    Clay's official problem statement accepts a proof of any one of four statements, two domains (all of space, or a periodic box) times two outcomes (with no external force, smooth solutions exist for all time; with some smooth force, a smooth solution breaks down). OpenAI proved a breakdown with a force. Most experts mean the unforced question by "the Navier–Stokes problem," and Constantin, Ignatova and Vicol showed that a construction with its key features needs a force that is neither analytic nor zero near the singularity, so such a construction cannot settle the unforced case. Luis Silvestre told Scientific American, "The Clay problem is settled, but the main problem for the Navier-Stokes equations is not."

    ↩
  2. N · 02

    Plasma marks each counted record accepted and full under its reading of the numbered question. Our audit checks attribution and prior recorded solutions, rather than independently refereeing the proofs. The latest case, Problem 1017, gives a uniform additive asymptotic estimate for clique partitions. Its formalization manifest explicitly distinguishes this from Formal Conjectures’ separate lower- and upper-bound targets. Dates describe the catalog’s publication record, with documented overrides; they are not dates of discovery or independent acceptance.

    ↩
  3. N · 03

    A kernel's verdict is only as good as the kernel, the axioms and the formal statement. Milo Moses's "A critical look at Lean 4" (October 2026) collects the weak points. Lean's kernel has had soundness bugs, which its creator Leonardo de Moura now says "AIs are really good at exploiting." Lean's logic has no published consistency proof relative to set theory: "As of writing, there is no publicly available proof that Lean4 is consistent with respect to ZFC + any large cardinal axiom." In a comment, Moses judges that the larger risk in practice is "statement mistranslation, not soundness issues."

    ↩
  4. N · 04

    Any human as we know them today, with a contemporary lifespan and only limited brain–computer interfaces.

    ↩
  5. N · 05

    Yehuda Rav, "Why do we prove theorems?", Philosophia Mathematica 7 (1999): 5–41. PYTHIAGORA is introduced on p. 5; the quotations are on p. 6.

    ↩
  6. N · 06

    I like to pretend mathematics is a principled grasp of structure above a dangerously unprincipled sea of formalism.

    ↩
  7. N · 07

    At current and falling inference costs, any mathematical question someone will pay to ask an AI model has economic value to its provider. That potential market spans all of mathematics, before we even consider its applications.

    ↩
  8. N · 08

    Even this is generous: much of primary and secondary mathematics education concerns itself with times tables, calculations and algebraic manipulations.

    ↩
  9. N · 09

    I develop this view at length in the paper on rigor and acceptance I mentioned at the start. I would welcome readers and feedback: reach me here.

    ↩
  10. N · 10

    Whether that reality is out there to be discovered is itself contested. I am more a structuralist than a pure Platonist, but you can make up your own mind.

    ↩