Partly, and only on specific kinds of task. AI systems can now give correct final answers to some competition mathematics, write ordinary-language proofs that people must judge, and produce formal proofs that a proof assistant checks. None of these results, by itself, shows that AI can carry out open-ended mathematical research, explain why a result matters, or take over the community work that mathematicians do. “Solving math” is several different tasks, and each one needs its own evidence.
Four tasks under one headline
The word “math” covers tasks that differ in what the system produces and in how anyone can tell whether the output is right.
As an Amazon Associate I earn from qualifying purchases.
| Task | What the system produces | How correctness is judged | Where it appears |
|---|---|---|---|
| Answer production | A number, expression or other final answer | Exact match to a known answer | IMO-Bench lists answer accuracy as one of its four dimensions (IMO-Bench project page) |
| Natural-language proof writing | A written argument that a reader must follow | Expert human evaluation; IMO-Bench states this remains the gold standard | IMO-Bench proof-writing dimension |
| Proof grading | A judgment on whether a given proof is correct | Not stated on the IMO-Bench project page | IMO-Bench grading dimension |
| Formal theorem proving | A proof written in a formal language such as Lean | A proof assistant checks it against formal definitions and rules | 2024 olympiad problems handled by AlphaProof and AlphaGeometry 2 (see below); IMO-Bench Lean proof tasks |
A system can score well in one row and tell you little about the others.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsThe 2024 olympiad result, read with its conditions
The clearest published data point comes from the Stanford Institute for Human-Centered AI’s Artificial Intelligence Index Report 2025. It reports that DeepMind’s AlphaProof and AlphaGeometry 2 solved four of six problems from the 2024 International Mathematical Olympiad at a silver-medal-equivalent level. The result is narrow: two systems, one competition, one year’s problem set. It is not a general measure of mathematical ability.
#1 Best Overall
Conditions that shape the claim
- The problems were translated by people first. The report notes that the problems had been manually translated into Lean, so the systems worked on formal versions prepared by hand.
- “Silver-medal-equivalent” is a score comparison. It describes a score level, not an official medal awarded to a system.
- Contest problems have known solutions. They test a bounded set of skills; they are not open questions that the field is still trying to settle.
- Standard theorem-proving benchmarks were not covered. The report said it remained unknown how these systems would perform on traditional theorem-proving benchmarks.
The result says nothing about whether such a system would find a new theorem, explain a proof to a student, or judge which open problems are worth pursuing. Those questions need different evidence.
What Lean checks, and what it cannot
Lean is a programming language and proof assistant in which a theorem and its proof are both written as code. Lean’s small trusted core, called the kernel, accepts a proof only if every step follows from the formal definitions and rules. A minimal example:
Rank #2
theorem add_swap (a b : Nat) : a + b = b + a :=
Nat.add_comm a bIf the proof does not fit the statement, Lean reports an error rather than accepting it. Running #print axioms add_swap lists the axioms the proof depends on. A proof that still contains a sorry placeholder will show sorryAx in that list, which makes it a quick check for an unfinished argument.
What a checked proof does not guarantee
- The statement must be the intended one. The checker confirms a proof of the formal statement. If that statement is weaker than, or different from, the question a mathematician meant, the checked proof answers the wrong question.
- Verification is not significance. A checked proof can be trivial, hard to read, or unhelpful for further work. Verification does not show that a result matters or is illuminating.
- The checker is software. A checked proof is only as reliable as the checker and the definitions it relies on. The kernel is kept small so that it can be inspected, but it is still code.
How IMO-Bench separates the skills
IMO-Bench, a project from Google DeepMind, is organised around four dimensions: answer accuracy, proof writing, grading, and Lean proof tasks. Its project page states that human expert evaluation remains the gold standard for mathematical proofs.
Separating these dimensions keeps one headline score from standing in for all four skills. A system can be accurate on final answers yet weak at writing proofs that experts accept, and a system that grades proofs well is doing a different job from one that writes them.
Who checks AI-generated mathematics
Two kinds of check apply. A proof assistant checks a formal artifact. Mathematicians judge whether a result is sound, clearly argued and important enough to build on. The first check can be automated; public material on this subject does not describe an automated equivalent of the second.
Rank #4
OpenAI made two announcements in September and October 2026 that bear on this question. Both are company statements, and the table separates what each says from what remains independently established.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Date | Source | What the company says | Status |
|---|---|---|---|
| September 21, 2026 | Advisory Group on Mathematics and Artificial Intelligence | Describes an advisory group whose remit includes review and communication of emerging results and academic and professional standards | The company’s description of the group’s role. Public material does not yet describe the group’s findings. |
| October 6, 2026 | Sharing AI progress in mathematics | Says the company is sharing Lean formalizations of many proofs and consulting an independent advisory group about release practices | A company statement. It is not independent verification of the proofs. |
Publishing Lean formalizations gives readers a formal artifact that a proof assistant can check. Whether the surrounding claims hold up is a separate question, and the mathematical community answers it.
Best Value
Earlier formal-math work
Formal-math AI predates the current headlines. In a 2022 post, OpenAI described a Lean theorem prover solving selected high-school olympiad problems (OpenAI, “Solving (some) formal math olympiad problems”). It shows that the formal approach has been developing for several years. It is not a current performance measure.
What remains unsettled
Public sources as of October 2026 do not establish which recent AI-generated mathematical claims have been independently reviewed and accepted by the wider mathematical community. Until that is clearer, treat each new claim as a report with its own task, conditions and date.
Quick Recap
A checklist for reading any AI math claim
- Task. Was the output a final answer, a natural-language proof, a proof grade, or a formal proof?
- Assistance. Were tools, internet access, human steering or extra time allowed?
- Problems. Were the problems public, held out, or translated into a formal language by experts?
- Judge. Was correctness set by an exact answer, expert review, or a proof assistant?
- Scope. Does the claim establish correctness only, or also explanation, novelty and research value?
- Attribution and date. Is the result company-reported or independently assessed, and when was it run?
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
