In 2023, AI systems scored essentially zero on International Mathematical Olympiad problems. In 2024, DeepMind's AlphaProof reached silver-medal level. In 2025, Google's Gemini Deep Think was reported to reach gold. That is a genuinely remarkable three-year climb — and it raises an honest question for every student and parent: if AI can now solve olympiad-level maths, what's still worth learning yourself? This article gives a straight answer, based on what's actually been demonstrated as of 2 October 2026, not on hype in either direction.
Not all "AI doing maths" is the same thing
"AI" covers several genuinely different kinds of tool, and conflating them is where most confident-sounding claims — in both directions — go wrong.
| Tool | What it actually does |
|---|---|
| Calculator | Exact arithmetic on numbers you give it. No reasoning, no mistakes within its design limits. |
| Computer algebra system (e.g. Mathematica, SymPy) | Exact symbolic manipulation — differentiation, factorising, solving equations — following fixed, verified algorithms. |
| Numerical software | Approximates answers that often have no exact form (e.g. simulating fluid flow), using well-understood error bounds. |
| Plain language model | Generates a step-by-step answer from learned patterns in text. Reasons fluently; can still make arithmetic or logical slips a calculator never would. |
| Language model with tools | Writes out its reasoning, then hands any actual calculation to a real calculator or code interpreter — combines fluent explanation with exact computation. |
| Formal proof assistant (e.g. Lean, Coq) | Checks every logical step of a proof against strict formal rules. If it accepts a proof, that step genuinely cannot contain a logical gap. |
| Specialised maths-reasoning AI (e.g. AlphaProof, AlphaGeometry) | Purpose-built and trained specifically for competition-style proof and geometry problems, often paired with a formal checker. |
The distinction students often miss
A plain language model guessing an answer is not the same as a formal proof assistant checking one
What olympiad gold does and doesn't tell you
Exam tipDeepMind's AlphaProof, paired with AlphaGeometry 2, solved four of six IMO 2024 problems for a silver-medal-range score — a huge jump from zero the year before. By 2025, Google's Gemini Deep Think was reported to reach gold-medal performance, solving problems entirely in natural language within the competition's own 4.5 hour limit. That's a genuinely impressive result on a genuinely hard benchmark: IMO problems demand real creative insight, not just calculation.
The distinction students often miss
Olympiad performance, research mathematics, and your homework are three different tests
Olympiad problems are short, famously difficult, but precisely stated, with a known correct answer to check against. Open research mathematics — like the Millennium Prize Problems covered in our companion article — has no known answer to check against at all, and typically demands building genuinely new mathematical structures, not just solving a well-posed puzzle. Your own coursework sits in neither category: it tests whether you can reliably reproduce a method, under exam conditions, with no tool at all. Strength on one of these three says comparatively little about the other two.
Where AI genuinely helps you learn maths
- Computation with the right tool. Asking an AI system that uses a real calculator or code interpreter for exact arithmetic or algebra is reliable, and faster than doing it by hand for checking purposes.
- Exploring patterns. Generating several worked examples of a technique quickly, to spot the underlying pattern yourself, before you formalise it.
- Suggesting a strategy. Asking "what approach would you try first?" on a problem you're stuck on, without asking for the full solution.
- Checking your own steps. Working a problem yourself first, then asking an AI to check specifically your method, not to redo it from scratch.
- Supporting revision. Generating extra practice questions on a specific weak topic, or explaining a concept a different way when a textbook's explanation isn't landing.
Where it still goes wrong — three real patterns to know
The three worked examples below are illustrative, constructed specifically to show these failure patterns clearly — not transcripts from a real AI system — but each one reflects a documented, well-known category of mistake.
1. Dropped domain restrictions
Exam-style question
Simplify (x² − 1) / (x − 1). [2 marks]
Show model answer and mark breakdown
- Factorise the numerator: x² − 1 = (x − 1)(x + 1) (1 mark)
- Cancel the shared (x − 1) factor: (x − 1)(x + 1) / (x − 1) = x + 1, provided x ≠ 1 (1 mark)
Examiner focus: the simplified form x + 1 and the original expression are not identical functions — at x = 1, the original is undefined (0/0), while x + 1 equals 2. The correct answer must keep the restriction: x + 1, x ≠ 1.
Common mistake: A fluent-sounding but incomplete answer: 'x + 1' alone, silently losing the x ≠ 1 restriction. This is an extremely common AI (and student) slip — the algebra is correct, but the simplified expression technically describes a slightly different function from the one in the question unless the restriction is stated.
2. Overgeneralising from a run of correct examples
The expression n² + n + 41 produces a prime number for every whole number n from 0 all the way to 39 — forty consecutive primes. That streak looks like overwhelming evidence of a pattern. Try the slider below.
Try it yourself — does n² + n + 41 always give a prime?
n = 10
| n | n² + n + 41 | Prime? |
|---|---|---|
| 6 | 83 | Yes |
| 7 | 97 | Yes |
| 8 | 113 | Yes |
| 9 | 131 | Yes |
| 10 | 151 | Yes |
| 11 | 173 | Yes |
| 12 | 197 | Yes |
| 13 | 223 | Yes |
| 14 | 251 | Yes |
n² + n + 41 (a formula studied by Euler) gives a prime for every n from 0 to 39 — forty straight primes, which looks overwhelming. Drag the slider to n = 40: the formula gives 1681 = 41², not prime at all. Forty correct examples in a row still isn't a proof, and the forty-first case is exactly where it breaks.
An AI system (or a student) extrapolating from n = 0 to 39 alone would have every reason to guess this always produces a prime — and would be wrong at n = 40.
This is exactly the kind of mistake a system reasoning from pattern alone — rather than proving a general rule — is prone to, whether that system is an AI or a person. Forty correct cases feel conclusive; they aren't.
3. Applying a rule without the understanding behind it
Our Derivatives article makes this point about students, and it applies just as directly to AI output: mechanically applying the power rule to differentiate x³ correctly gives 3x² — but that same mechanical pattern-matching, applied to a function built from several rules chained together (a product of two functions, say, or a function inside another function), can silently apply the wrong rule, or apply the right rule to the wrong piece, and still produce a confident, fluent-looking wrong answer. A system (or a student) that understands why the chain rule exists — change rippling through connected quantities — can sanity-check whether an answer makes sense. One that has only pattern-matched the mechanical steps cannot.
The distinction students often miss
This is not "AI can only pattern-match, full stop"
Check your understanding
An AI tool simplifies (x² − 1)/(x − 1) and writes just 'x + 1', with no restriction stated. What's the most useful next step?
A practical workflow
Used this way, AI strengthens your maths rather than replacing it:
- Attempt the problem yourself first, even partially — struggling productively is part of how understanding forms.
- Ask for a hint, not the answer, if you're stuck — "what should I try first?" rather than "solve this."
- Explain the method back in your own words — if you can't, you don't own it yet, regardless of where it came from.
- Check assumptions and domain restrictions explicitly — the single most common place a fluent-looking answer quietly goes wrong.
- Verify independently — a different method, a sanity-check estimate, or a known special case.
- Try a closely related problem with no assistance at all — the real test of whether the understanding transferred.
Where to go next
If the idea of "evidence versus proof" from the n² + n + 41 example interested you, our Millennium Prize Problems article and the full Unresolved Mathematical Mysteries hub are full of real, currently-open examples of exactly this gap, at the very top of the field.
Sources and further reading
- AI achieves silver-medal standard solving International Mathematical Olympiad problems — Google DeepMind — AlphaProof + AlphaGeometry 2, IMO 2024
- Olympiad-level formal mathematical reasoning with reinforcement learning — Nature
- From Silver to Gold: A Technical Look into AI's Conquest of the IMO — 2025 gold-medal-level results
- Live from ICM 2026: What Is Math For in the Age of AI? — Quanta Magazine
Last reviewed: 2 October 2026. Figures are typical published values and rules can change. Exam wording varies by specification, so check yours.
Frequently asked questions
So is it safe to just let AI do my maths homework?
For checking your own working, exploring 'what if' variations, or getting unstuck on one step, yes — used that way it's a genuinely useful study tool. For producing an answer you then hand in without understanding it, no: you lose the exam-room skill (no AI allowed there), and AI reasoning — however good — is still wrong often enough, in ways that aren't always obvious, that unverified output is a real risk.
Didn't AI already win gold at the International Mathematical Olympiad?
Yes. Google's Gemini Deep Think was reported to achieve a gold-medal-equivalent score at IMO 2025, working entirely in natural language within the standard time limit — a genuinely major result, one year after DeepMind's AlphaProof reached silver-medal level at IMO 2024. Olympiad problems are a specific, famously difficult but well-defined kind of maths, though — strong olympiad performance doesn't automatically transfer to every kind of mathematical work (see "What olympiad gold does and doesn't tell you" below).
Why does AI sometimes get simple maths wrong if it can solve olympiad problems?
A general-purpose language model generates its answer token by token from learned patterns, which is a different process from mechanically executing arithmetic — it's why the same system can produce an elegant, correct olympiad-style argument and still slip on a multi-digit multiplication in the same conversation. Systems that hand arithmetic off to an actual calculator or code interpreter avoid this specific failure; a 'pure' language model answering from text alone does not.
Will understanding maths still matter once AI gets even better?
If anything, more, not less. The more routine calculation gets automated, the more the genuinely human skill — knowing which question to ask, spotting when an AI-confident answer is actually wrong, and understanding a result well enough to apply it to a new situation — becomes the part nobody can outsource. That's precisely the argument our Derivatives article makes about learning the 'why' before the 'how', and it applies here even more directly.
Reader ratings & reviews
Rate this article across a few different things — your rating helps other readers, and helps us improve.
Was this article helpful?
Tell us what you think — a correction, a question it left unanswered, or a topic you'd like covered next.
Prefer To Talk? Book a free video call at a time that suits you: 15 minutes for a quick question, 30 for anything else. Times show in your own time zone.
About the author
Sudershan Soni
Founder & Lead Tutor at Mostak Services — an MSc-qualified Mathematics, Science, Computer Science & STEM tutor with 20+ years of professional experience, teaching students from 11+ and GCSE to A-Level and beyond, online worldwide.
Read full profile