Computers can't do original mathematical reasoning; no machine could win a math olympiad
The old claim turned out to be wrong and was rejected by evidence.
TaughtComputers are fast calculators, but genuine mathematical creativity (reading a hard, unseen olympiad problem and inventing a multi-step proof) is a distinctly human act. A machine can crunch numbers; it cannot reason its way to a novel proof under exam conditions.
NowIn July 2025, AI systems from Google DeepMind and OpenAI both scored 35 out of 42 on the International Mathematical Olympiad, gold-medal level, solving five of six problems as free-form written proofs, under the same rules and time limits as human contestants.
What actually happened
The IMO is not arithmetic. Its problems are short to state and brutal to solve, demanding the kind of inventive, several-page argument that had been a reliable line between calculation and understanding. As recently as 2024, the best systems reached silver and needed problems hand-translated into formal logic. A year later they read the problems in ordinary language, wrote ordinary proofs, and cleared the gold threshold. DeepMind's result was officially graded by the IMO.
The honest boundaries: this is not a proof that machines 'understand' mathematics the way a mathematician does, they still failed the hardest problem, and one strong result is not a general guarantee. But the specific, confident claim taught for decades, that composing original proofs to unseen competition problems was beyond any computer, did not survive 2025. The line moved, and fast.
Newsletter
Get new weekly revisions
Sources
Who was taught this
Still standard through 2025, so anyone who finished school between 1950 and 2025 learned the earlier version.