AI is now doing real mathematics

In May 2026, AI from OpenAI and Google DeepMind produced original results on long-open Erdős problems — questions easy to state, impossible to fake, with no memorised answer to copy. OpenAI's model disproved a geometry conjecture dating to 1946. DeepMind's AlphaProof Nexus solved nine open Erdős problems plus 44 conjectures from the integer-sequence catalogue, two of which had been unsolved for 56 years. Different companies, different methods, same week.

DeepMind's results are verified step-by-step in the Lean proof assistant — the AI proposes, Lean certifies, and a flawed step is simply rejected and fed back. That guards against AI's worst habit: sounding right while being wrong. OpenAI's construction was checked by human mathematicians, including Timothy Gowers, with formal peer review still to come — credible, but a softer guarantee. Same word, 'solved'; two different levels of proof.

DeepMind's own boss, Demis Hassabis, called it "still not AGI", WinBuzzer reported: useful for maths that can be checked, not a thinking machine (WinBuzzer).

The durable result isn't any single proof — it's the pipeline. 'AI proposes, formal checker certifies' neutralises AI's worst failure mode wherever claims can be mechanically verified. Mathematics was the hardest proving ground precisely because the bar for correctness is absolute: Lean either accepts every step or it doesn't. If that generate-then-verify template spreads to software testing, computational biology, or other formalisable fields, the May 2026 maths results will look, in hindsight, like the prototype — the moment a reliable pattern was first demonstrated — rather than the prize itself.