November is fully booked! Next classes blast off the week of Mon, Dec 7.

Reserve your spot
Verifying an AI Maths Proof: The Navier-Stokes Case
Table of Contents

Verifying an AI Maths Proof: The Navier-Stokes Case

Verifying an AI maths proof takes six gates, not one. How mathematicians will check the September 2026 Navier-Stokes claim, and how long each step takes.

Verifying an AI maths proof is not one event. It is six, and the computer only handles the first. OpenAI’s September 8, 2026 claim about the Navier–Stokes problem came with a Lean formalization, and OpenAI reported an additional 17 hours of verification work via GPT-6 Astra. That closes gate one. Wikipedia’s entry on the Millennium Prize Problems records that as of October 2026 the solution “awaits scientific validation,” which is gate six.

The five gates in between are human, and they are the interesting part. Knowing what they are is the difference between reading the news and understanding it.

Key Takeaways

  • A proof assistant like Lean “checks that these proofs are correct down to their logical foundation,” per the Lean community. That rules out logical gaps in the argument as written. It rules out nothing else.
  • The gate nobody can automate is statement fidelity: whether the formal theorem in the file is actually the Clay problem, as written by Charles L. Fefferman in the official problem description.
  • Formalization audits are routine and specific. A reviewer checks whether the file used placeholder steps, extra axioms, or a modified library, and the proof assistant will report which axioms a theorem depended on.
  • Clay’s process is slow by design: the Institute “only considers proposed solutions to Millennium Prize problems after at least two years since publication,” and Martin Bridson, its president, emphasised peer-review requirements.
  • Named human graders remain the standard even for machine work. The 2024 International Mathematical Olympiad results from Google DeepMind were scored by Prof Sir Timothy Gowers and Dr Joseph Myers under official rules.

Verifying an AI Maths Proof: Six Gates, in Order

Gate 1. Does the file type-check?

Lean is a proof assistant, which the community site defines as “a piece of software that provides a language for defining objects, specifying properties of these objects, and proving that these specifications hold.” If the proof term type-checks, every inference step is valid in Lean’s logic. This is mechanical, fast relative to human review, and extremely strong within its scope. OpenAI’s reported 17 hours of additional verification falls here.

Gate 2. Does the formal statement match the real problem?

This is the gate that matters most and the one a machine cannot touch. Somebody fluent in both Lean and partial differential equations has to read the theorem statement in the file and compare it, line by line, against Fefferman’s official formulation on the Clay site. The question is narrow and brutal: is this the same claim?

Getting this wrong is not exotic. A formalisation can prove something true, important and adjacent while missing the actual statement by one quantifier. Nobody detects that by running the checker, because the checker has no opinion about which theorem you should have stated.

Gate 3. Are the definitions standard?

Navier–Stokes verification depends on the precise meaning of several terms: which function space the solution lives in, what “smooth” means here, what regularity the external force is required to have. Fefferman’s statement fixes these. A reviewer checks whether the formalisation’s definitions are the standard ones from mathlib or custom definitions written for this proof. Custom is not automatically wrong. It is automatically something to read very carefully.

Gate 4. What did the proof actually assume?

Formalization has a specific audit trail. A reviewer can ask the system to report which axioms a given theorem depended on, which catches three things at once: whether any step was left as a placeholder, whether extra axioms were introduced beyond the standard foundation, and whether the library itself was modified. A proof that compiles against a patched library is a different object from one that compiles against the public one.

Gate 5. Peer review.

A refereed mathematics journal assigns specialists who read for correctness, significance and whether the result is what the abstract says. For a Millennium problem this takes a long time and involves several people who will not be in a hurry. Clay’s rules make publication in a refereed venue a prerequisite.

Gate 6. Community acceptance.

The last gate has no form to fill in. It is reached when specialists in the field stop objecting and start building on the result. Clay then waits at least two years from publication before it will even consider a submission. The Institute’s own public response to the September 2026 announcement was that it “shares in the excitement of the global mathematical community as we contemplate the announcement,” which is a sentence engineered to commit to nothing.

Now the analogy, with the mechanism already in place. Gate 1 is a spell-checker: it catches every misspelling and has no view on whether you wrote the right essay. Gate 2 is reading the essay question. Gates 3 and 4 are checking you did not quietly redefine the words. Gates 5 and 6 are the examiners, and the examiners take their time.

Who Does the Checking, and How Long Each Gate Takes

GateWho does itRough timescaleWhat it rules out
1. Type-checkThe proof assistantHours to daysLogical gaps in the argument as written
2. Statement fidelityA specialist fluent in both Lean and PDEWeeksProving the wrong theorem
3. Definition auditPDE specialistsWeeksNon-standard definitions changing the meaning
4. Axiom and library auditFormalization specialistsDays to weeksPlaceholders, extra axioms, patched libraries
5. Peer reviewJournal refereesMonths to yearsErrors, overstated significance, unclear exposition
6. Community acceptanceThe fieldYearsEverything the first five missed

Notice the asymmetry. The computational gate is the fastest and the narrowest. The slow gates are the ones about meaning, and meaning is where disputes live.

Two names are worth knowing in this particular case. Tristan Buckmaster of NYU, a specialist in exactly this area, alleged that OpenAI used non-public results of his. OpenAI initially said it “cannot rule out that de-identified data” from his use of its products “helped improve our models,” and later said an investigation concluded his prompts “could not have influenced the system.” Levent Alpöge, a mathematician at Anthropic, was also named in connection with work allegedly used. These are allegations and company statements, not findings, and a careful reader keeps that distinction.

How to Teach Your Kid About Verifying a Proof

Ages 5–8: Find the broken step

Materials: paper and pencil.

Write out a chain of five simple steps with exactly one error. For example: 2 + 3 = 5, 5 + 4 = 9, 9 + 1 = 11, 11 + 2 = 13, 13 + 3 = 16. The third step is wrong.

Your child’s job is to find it. Most children check the last line first and declare the whole thing wrong. Redirect: “check each step, in order, and stop at the first one that breaks.” That is exactly how a proof checker works, and the habit is the lesson.

Then the second round, which is better. Make the chain arrive at a correct final answer despite containing two errors that cancel out. Now the right answer hides a broken argument. Ask which matters more, the answer or the steps. In mathematics it is always the steps.

Ages 9–12: The assumption hunt

Materials: index cards.

Write three short arguments, each with a hidden assumption. “Every bird I have seen can fly, so all birds can fly.” “This shape has four sides, so it is a square.” “The light turned on when I flipped the switch, so the switch controls the light.”

Your child finds the assumption in each one and writes it on the back of the card. Penguins exist. Four sides does not mean equal sides or right angles. Something else might have turned the light on.

Then flip the exercise. They write a three-step argument about something in the house, and you find the assumption. This is gate 2 and gate 3 in miniature, and nine-year-olds are often better at it than adults because they are less polite about it.

Ages 13+: Make the checker complain

If your teenager is willing to try Lean through its free web interface, this is the single most instructive twenty minutes available.

Have them write a theorem statement and leave the proof as a placeholder instead of proving it. The system accepts the file and flags that the declaration relied on a placeholder. Then have them ask the system which axioms the theorem depended on, and watch the placeholder appear in the list.

That is gate 4, demonstrated. The point is not that they learn Lean. The point is that they see a system that distinguishes between “this file compiles” and “this theorem is proved,” because that distinction is the entire reason anybody trusts formalisation.

The question to ask: “If the computer says the proof is correct, what is left for a human to check?”

What to Do at Home

Watch for the Lean file, not the announcement

A claim with a public, downloadable formalisation that independent people can re-verify is in a different evidentiary class from a company blog post. When a kid asks whether the Navier–Stokes result is real, the honest answer is to look for the file and the people reporting that they ran it. Teaching that instinct is more valuable than any conclusion about this particular claim.

Use “which theorem did they prove?” as the default question

For every machine-assisted mathematics result that makes the news, ask this one question out loud. It is gate 2, it is where most of the ambiguity lives, and a thirteen-year-old can ask it without knowing any of the mathematics. The answer for September 2026 is specific: statements (C) and (D) of the Clay problem, which assert breakdown rather than smoothness.

Separate speed from certainty

OpenAI described finishing the work in around 88 hours, with roughly 10,000 concurrent agents, 2.7 million messages and about 130 billion output tokens. That is astonishing speed. Verification will take years. Both facts are true at once, and holding them together is a genuinely useful habit of mind.

Read a journal’s instructions for referees

Most journals publish them. They are short, dry and oddly clarifying, because they show exactly what a referee is asked to decide: correctness, novelty, significance, clarity. Reading one with a teenager demystifies peer review better than any explanation of it, and makes the two-year Clay rule feel like a reasonable precaution rather than bureaucracy.

What not to do

Do not conflate “unverified” with “false.” The September 2026 claim may well hold. It has a formalisation, which is more than most historic claims on famous problems ever had. The correct posture is suspended judgement, and modelling suspended judgement in front of a kid is harder and more valuable than modelling either enthusiasm or scorn.

What to Watch For Over the Next 3 Months

  • Week 4: Check for an arXiv preprint and a public repository containing the Lean development. Also check whether any named mathematician has reported independently type-checking the file. Independent re-verification closing gate 1 outside the claiming company is the first real external milestone.
  • Month 2 red flags: Objections aimed at the formal statement rather than the proof steps. That would be a gate 2 problem, and it is the serious kind. Also watch for disclosure that the proof used custom definitions or a modified library, which is a gate 3 and gate 4 question rather than a scandal.
  • Month 3 self-check: Ask your teenager which gate the claim is currently stuck at and why a computer cannot close it. If they say gate 2 and explain statement fidelity, they understand the story better than most of the coverage. If they say “it needs more computing,” revisit the six gates.

Frequently Asked Questions

If Lean verified it, is the proof correct?

The argument is logically valid within Lean’s foundation, which is a strong guarantee. Whether the theorem proved is the Clay problem, whether the definitions are standard, and whether placeholders or extra axioms were used are separate checks. Correct and relevant are different properties, and only the first is mechanical.

How long will verification take?

Nobody can give a date, but the structure sets a floor. Clay requires peer-reviewed publication and then “at least two years since publication” before it will consider a submission. Gate 2 alone has historically taken weeks to months on far less complicated claims.

Who is qualified to check this?

A small group: specialists in the regularity theory of fluid equations, and specialists in Lean formalization. Very few people sit comfortably in both groups, which is a real bottleneck. The International Mathematical Olympiad precedent is instructive: Google DeepMind’s 2024 results were graded by Prof Sir Timothy Gowers and Dr Joseph Myers, both named, both human.

What is the most common way a formalisation goes wrong?

Proving the wrong statement. The logic can be flawless while the theorem differs subtly from the intended problem, often by a quantifier, a function space or a regularity assumption. This is why gate 2 is read line by line rather than skimmed.

Does the priority dispute affect whether the proof is valid?

No. Validity and attribution are separate questions. A proof can be correct and still raise legitimate questions about whose unpublished ideas it used. Keeping those two threads apart is good practice, and it is also fair to the mathematicians involved on all sides.

Should my kid learn formalisation or classical proof writing?

Both, starting with classical. Writing proofs by hand builds the judgement that gates 2 and 3 require, namely whether a statement captures what you meant. Formalisation then teaches precision and makes hand-waving impossible. The two reinforce each other, and the order matters.


About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.


Sources

  1. OpenAI. (2026, September 8). “Navier–Stokes solution.” https://openai.com/index/navier-stokes-solution/
  2. Wikipedia contributors. (2026). “Millennium Prize Problems.” Wikipedia. https://en.wikipedia.org/wiki/Millennium_Prize_Problems
  3. Clay Mathematics Institute. “Navier–Stokes Equation,” official problem description by Charles L. Fefferman. https://www.claymath.org/millennium/navier-stokes-equation/
  4. Lean Community. “Lean and mathlib.” https://leanprover-community.github.io/
  5. Google DeepMind. (2024, July 25). “AI achieves silver-medal standard solving International Mathematical Olympiad problems.” https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/
  6. Wikipedia contributors. “Navier–Stokes existence and smoothness.” Wikipedia. https://en.wikipedia.org/wiki/Navier%E2%80%93Stokes_existence_and_smoothness
  7. Clay Mathematics Institute. “Millennium Problems.” https://www.claymath.org/millennium-problems/

Related reading on HiWave Makers: how mathematicians check a 125-page AI proof, whether AI can do original mathematics, and proof assistants your kid can try.

Ricky Flores
Written by Ricky Flores

Founder of HiWave Makers and electrical engineer with 15+ years working on projects with Apple, Samsung, Texas Instruments, and other Fortune 500 companies. He writes about how kids learn to build, think, and create in a tech-driven world.