AI Has Started Producing Serious Mathematics. ‘Solved’ Is Still the Wrong Default Word

It is increasingly common to hear that ChatGPT has “solved” difficult mathematical problems. That sentence captures something real, but it also puts several very different achievements into one box.

The real development is more interesting than the slogan. AI systems are beginning to generate research-grade mathematical artifacts: long proof attempts, computer-checkable formalizations, code, revisions prompted by critics, and public repositories that other mathematicians can inspect. In recent releases, OpenAI has described an internal system working on the First Proof challenge, a claimed result on a version of the Navier–Stokes problem, and a broader archive of mathematical results with accompanying Lean files.

That is a major shift from asking a chatbot to explain calculus. It is not, however, a licence to treat every headline as a theorem accepted by the mathematical community.

The distinction matters for readers, researchers and anyone building with AI. A candidate idea is not yet a proof. A proof written in prose is not yet a proof checked by a formal system. And a formal proof of a precisely encoded statement is not automatically the same thing as an independently reviewed conclusion about the informal problem a headline describes. These stages reinforce one another; they are not interchangeable.

The important correction: it is not just “ChatGPT answering a question”

The phrase “ChatGPT solved it” can obscure the kind of system being discussed. OpenAI’s September release on Navier–Stokes describes an internal model, not an ordinary one-shot interaction in the consumer ChatGPT interface. According to OpenAI, groups of agents could read a cached web corpus, run code, exchange intermediate results and be reorganised as the search progressed. The company says the group that generated its claimed Navier–Stokes resolution involved roughly 10,000 concurrent agents.

That is closer to a computational research operation than to a familiar chat session. It includes task decomposition, parallel search, critique, synthesis, tool use and, afterward, formalisation in Lean. The reported scale does not prove the mathematics by itself. It does explain why comparing the result with a single model response is misleading.

This broader pattern is visible outside OpenAI too. Google DeepMind’s 2024 AlphaProof and AlphaGeometry 2 work was presented as a specialised system that tackled International Mathematical Olympiad problems with formalised inputs and generated proofs, reaching a silver-medal-level score in the company’s account. The emerging capability is therefore not simply fluent mathematical language. It is a pipeline that can search mathematical space while using representations that machines can inspect.

First Proof shows why revision belongs in the headline

The strongest reason to be careful comes from OpenAI’s own First Proof update. First Proof was designed around research-level problems that require complete arguments in specialised areas, where correctness cannot be quickly settled without experts. OpenAI ran an internal model on all ten problems and, after feedback, said that five attempts had a high chance of being correct while several others remained under review.

That statement is notable. So is what followed: OpenAI said it had initially thought its attempt on problem 2 was likely correct, then changed its position after the official commentary and further community analysis.

This is not an embarrassment to hide in a footnote. It is the most useful lesson in the release. Mathematical research is designed to make attractive errors survivable. A proposed proof may contain a hidden assumption, use a theorem outside its valid domain, or establish a narrower statement than readers think. Criticism, reworking and public checking are how a field distinguishes an intriguing argument from a durable result.

An AI system that can revise after criticism is more useful than one that merely sounds certain. But a system’s willingness to revise is also evidence that the first answer should not be called final too early.

What a Lean proof changes — and what it does not

Formal proof assistants such as Lean are important because they turn a proof into an object a computer can check against explicit definitions and logical rules. That can remove entire classes of unnoticed algebraic slips or omitted steps. If a formalization compiles without admitted gaps, the encoded theorem has unusually strong mechanical support.

OpenAI’s Navier–Stokes repository includes Lean formalizations and says the files contain no `sorry` declarations, a common marker for an unproved placeholder in Lean. That is meaningful evidence. It makes the claim more auditable than a PDF alone.

Still, “formalized” is not a magical synonym for “the public controversy is over.” There are two questions, not one. First: does the proof assistant verify the exact formal statement, given the selected definitions, axioms and dependencies? Second: does that formal statement match the scope and interpretation of the informal theorem described in the paper and the news coverage?

The second question can require domain expertise. A model may formalize a technically correct theorem about a forced variant of an equation, while a casual reader hears “the famous fluid problem is settled” in a broader sense. The repository’s own metadata is also useful context: its formalization review status is labelled self-assessed. That does not invalidate the files. It tells readers what kind of review has been documented so far.

Navier–Stokes: a claimed resolution, not a prize announcement

OpenAI says it has released an analytical proof and Lean formalization showing finite-time singularity formation for a forced three-dimensional Navier–Stokes setting, corresponding to specific alternatives in the official problem formulation. It calls this a resolution of the Millennium Prize problem. It also says explicitly that it does not intend to claim the Millennium Prize.

That last sentence should guide how the story is written. The Clay Mathematics Institute’s rules set a process involving publication, time and acceptance in the mathematical community before a prize can be considered. A company release—even an unusually detailed one with formal files—is not the same thing as completing that process.

The fair reading is neither dismissal nor premature coronation. If the argument withstands external scrutiny and its formalisation accurately captures the relevant problem, it could be an extraordinary research event. If expert review identifies a gap, the release may still be valuable: it exposes methods, conjectures and formal machinery that others can build on. The public should be able to hold both possibilities at once.

Why this is still a real turning point

Being precise does not mean being complacent. Until recently, the common limitation of general AI in advanced mathematics was not merely occasional arithmetic error. It was the inability to sustain a long, specialised chain of reasoning while managing definitions, cases, objections and dependencies.

OpenAI’s October mathematics release points toward a changed operating model. The company published a GitHub collection of results, many Lean formalizations, summaries of reasoning and estimates of compute. It also said it was consulting an independent advisory group and that the quality of citations, exposition and presentation still needed improvement. That combination—ambition joined to an explicit acknowledgement of unfinished scholarly practice—is more informative than a benchmark score.

The practical innovation is auditability. Researchers can ask: What was the precise statement? Which assumptions were used? Is there code? Is there a formal certificate? What changed after review? What remains self-assessed? Those questions transform “the model says so” into a research workflow with inspectable checkpoints.

This is the direction that could change everyday mathematical work first. A strong system may not replace a mathematician’s judgement about what problem matters or what abstraction is elegant. It may, however, generate alternative lemmas, test edge cases, search constructions, translate arguments into formal language, and make a large volume of proof work easier to interrogate.

The new bottleneck is evaluation

As systems become better at producing candidate results, the scarce resource may shift from generation to evaluation. Senior researchers will still be needed to decide whether an argument is interesting, whether its assumptions are natural, whether a formal theorem expresses the intended claim, and whether an explanation helps other humans extend the work.

There is also an access question. A search involving thousands of concurrent agents and very large token budgets is not an ordinary laboratory instrument. If only a few companies can run it, the mathematical community will need stronger norms around release, reproducibility, attribution and independent review. Publishing proof artifacts is a promising start, but it is not the end of that conversation.

For students and professionals, the practical rule is simple: use AI as a serious collaborator, never as the final authority. Ask it to expose assumptions, produce counterexamples, formalise a lemma or compare approaches. Then make the claim proportionate to the evidence.

What readers should watch next

Watch for external mathematicians engaging directly with the released papers and Lean files, not merely repeating company announcements. Watch for revisions that clarify a theorem’s scope, and for independent attempts to reproduce the formal builds. Watch, too, for whether companies make their evaluation protocols and artifacts easier for scholars without vast compute budgets to inspect.

The phrase “AI solved mathematics” is too blunt for what is happening. A more accurate statement is both humbler and more consequential: AI has begun to participate in the production of serious mathematical research, and the standards of proof, review and explanation now matter more—not less.

Signal & Syntax will continue to update this guide as products, access and practical evidence change.