Skip to main content
AI & Innovation

AI Math Proofs: What Google's Lean Results Prove

AI math proofs just hit 1.8 million machine-checked lines. See what Google's Lean results really prove, what they don't, and where humans still matter.

Nora Whitfield
November 22, 20269 min read
Share
AI Math Proofs: What Google's Lean Results Prove
Table of contents

Picture yourself as a math reviewer handed a 1.8 million line proof. Nobody on Earth could read it line by line. Yet in November, Google said its AI agents had produced exactly that, and that a computer had already checked every single step. AI math proofs have moved from contest puzzles to some of the deepest results in the field.

You don't need to follow the math to see why this matters. The way a proof earns trust is changing. For centuries that trust came from a handful of experts reading carefully. Now it can come from a small program that refuses to accept any step it cannot verify.

This piece explains what Google actually claimed, what "machine-checked" does and does not guarantee, and where humans still sit in the loop.

What Google Says Its AI Math Agents Did

The announcement came on the Google Research blog on November 13, written by Google Fellow Vahab Mirrokni. It describes /teamwork, a multi-agent feature inside Google's Antigravity development platform, built on its Gemini 4 models.

The system splits a big problem across many specialist agents. Some write proof steps. Others act as critics and auditors that check results without seeing how they were produced. If you've read our look at when adding more AI agents stops paying off, this is the same idea pushed into pure mathematics, with a strict referee attached.

Here are the math claims, as Google reported them:

  • Weil conjectures: an end-to-end formalization of about 1.8 million lines of Lean, built from 28,800 subagent invocations. Google calls this run "fully autonomous." Pierre Deligne finished the original proof in 1974, as Interesting Engineering noted in its coverage.
  • Hilbert's tenth problem: more than 919,000 lines of Lean, the first formalization of a recent proof that extends the result to general rings of integers.
  • Courtade-Kumar conjecture: a proof in full generality, formalized at roughly 50 million lines of Lean 4.
  • Bochner-Riesz in three dimensions: a formalization during which the agents found a gap in a human-written proof, which was then fixed.

Interesting Engineering's report repeated the Weil and chemistry figures and quoted no outside mathematicians. So every number above is Google's account of its own work. Keep that in mind as you read on.

Why a Lean Proof Checker Changes Trust

Lean is an open-source programming language and proof assistant. You write a theorem and its proof in a strict formal language. A small core program, called the kernel, then checks that each step follows from definitions and earlier steps.

The kernel does not care who wrote the proof. A human, a student, or 28,800 AI subagents all face the same test. That is why a falsifier like this suits an AI system so well. A large language model guesses, often wrongly, and the checker throws out every wrong guess automatically.

Google's post says this directly. Model output is unpredictable, so a single attempt may miss. The fix is to generate many candidates in parallel and let tools like the Lean checker, test suites and simulators reject the bad ones.

The part a checker cannot verify

A checker proves that your proof matches your statement. It cannot tell you whether your statement matches the problem you care about. If someone encodes a theorem slightly wrong, Lean will happily certify a proof of the wrong thing.

Google says its human collaborators confirmed that each formal statement matched the original problem. That step is where the real expertise now sits. Reviewing a few hundred lines of definitions is human work. Reviewing 1.8 million lines of proof is machine work.

A Worked Example: Why Nobody Reads These Proofs

Here is a rough way to feel the scale. Assume a skilled mathematician can carefully check 50 lines of formal proof per hour. That is generous for dense material.

  1. The Weil formalization is about 1,800,000 lines.
  2. At 50 lines an hour, that is 36,000 hours of review.
  3. At 2,000 working hours a year, that is 18 years for one person.
  4. The Courtade-Kumar file, at about 50 million lines, would take 500 years at the same pace.

So human line-by-line review is off the table. The only practical reviewer is the checker itself. That is fine, as long as you trust a small, well-tested kernel more than you trust a tired human on page 40,000. Most proof assistant users do, because the kernel is a tiny program that many people have examined for years.

What this changes is the work a mathematician does. Less time goes on reading. More goes on writing precise statements, and on checking that the definitions mean what the field means by them.

Formalizing a Proof vs Finding a New One

Two very different achievements are mixed together in the headlines. Keeping them apart helps you judge the news.

Formalization takes a proof humans already found and rewrites it in a language a computer can check. The Weil conjectures and Hilbert's tenth problem fall here. They are huge engineering jobs, because informal proofs skip steps that experts treat as obvious. A machine has to fill in every one of them.

Discovery means finding a proof nobody had before. Google says the Courtade-Kumar result belongs here. According to the blog, Gemini first found counterexamples to a broader version of the problem. The agents then reduced the conjecture to a 14-variable optimization problem and extended the proof from balanced to unbalanced functions.

A formalization tells you a known proof has no holes. A discovery tells you something new is true. The second is rarer, and it is the one that needs outside mathematicians to read the human-facing write-up.

The Bochner-Riesz case sits in between, and it may be the most useful result for working researchers. The agents found a gap in an informal proof. That is exactly the kind of error peer review misses, because reviewers rarely have time to rebuild every step.

AI Math Proofs Are Already Reaching Peer Review

The same toolkit includes a feature called Deep Review, which reads papers and flags problems. Google tested it on submissions to STOC 2027, a major theoretical computer science conference. Here is what the company reported:

  • Authors of 578 submissions opted in, which Google says was more than a third of all papers.
  • Of 70 authors who answered a survey, 90% rated the feedback mostly or almost all factual.
  • More than 30% said it caught substantive bugs in their proofs.
  • 14% said it found technical gaps that took over an hour to fix.

A survey of 70 people is small, and the people who opt in may like AI tools more than most. Still, the pattern is clear. AI is entering math first as a checker, not as a replacement mathematician. It catches errors faster than a reviewer with a deadline can.

This mirrors what happened in medicine. Our review of what AI drug discovery has actually found showed the early wins came from narrowing the search, not from replacing the scientist.

What Google Did Not Tell You

The blog post is detailed about results and thin on cost. It reports no compute budget and no dollar figure for these runs. Only a few time spans appear. A software rewrite took a two-week campaign, and one benchmark run lasted 11.5 hours.

That matters for you if you plan to use these tools. A result that took thousands of chip-hours is a proof of concept. A result that runs overnight on a modest budget is a new way of working. From the outside, nobody can yet tell which one this is.

Several items in the post are also described as under review or still in progress. Independent mathematicians have not yet published assessments of the Courtade-Kumar proof. Treat that one as a strong claim awaiting confirmation.

Who Gets Access, and What It Means for You

Google says the Gemini 4 models with /teamwork and Deep Review are now open to outside users inside Antigravity. It is also taking applications from 10,000 academic researchers for six months of free access to its highest usage tier. Deep Review gets two free calls a week.

If you work in research, the practical moves look like this:

  • Use the checker for what it is good at: verifying long, fiddly arguments you already believe.
  • Spend your own time on statements: check that every formal definition means what your field means.
  • Run review tools before submission: a reviewer will not thank you for a gap a machine could have flagged.
  • Ask about cost before planning around it: free access is temporary, and no pricing for heavy use has been published.

The same pattern also reaches beyond math. Google applied the toolkit to code, rewriting a video decoder into about 170,000 lines of Rust. It says the port came out 20% to 30% faster than the original C++ after tuning. Any algorithm with a precise test suite can be attacked the same way: generate, check, discard, repeat.

Frequently Asked Questions

Can AI prove math theorems on its own?

Partly. Google says its Weil conjectures formalization ran fully autonomously, but that proof was already known. For new results, humans still chose the problems, wrote or confirmed the formal statements, and edited the output for publication.

What is Lean and why do mathematicians use it?

Lean is an open-source proof assistant. You write definitions, theorems and proofs in a formal language, and a small kernel checks every step. Mathematicians use it because a proof that passes the checker cannot contain a hidden logical gap.

Are AI-generated proofs trustworthy?

A proof accepted by Lean is logically valid for the statement it proves. The remaining risk is a statement that does not match the intended problem. That is why human experts still need to review the definitions and theorem statements.

Will AI replace mathematicians?

Nothing in Google's results points that way. The work shifts toward framing problems and checking formal statements, while machines handle long verification. The Deep Review survey suggests AI is arriving as a reviewer first.

The Bottom Line on AI Math Proofs

Google's claims are big, but the most important shift is quieter. Trust in a proof is moving from who reviewed it to what checked it. A 1.8 million line proof no human can read is still trustworthy, if the statement is right and the kernel accepted it.

The open questions are cost and independent confirmation. Watch for outside mathematicians to weigh in on the Courtade-Kumar proof, and for Google to publish what these runs cost. Those two answers will tell you whether this is a showcase or the new normal for research.

Share

Written by

Nora Whitfield

Technology & Science

Covers technology and science for Quick Trend Insights, with a focus on what new AI tools, consumer hardware and space missions actually change for the people following them.

More articles

Related Articles

View all