OpenAI’s maths proof is a milestone, not a great leap
TL;DR:
- An internal OpenAI model disproved the Erdős unit distance conjecture, a problem in discrete geometry unsolved for around 80 years.
- Leading mathematicians, including Fields Medallist Tim Gowers, called it a genuine milestone for AI in mathematics.
- The result reused established ideas across several subfields rather than inventing new methods, and human mathematicians later refined it.
OpenAI’s announcement, made in mid-May, is arguably the first time an AI system has produced a proof resolving a major open conjecture on its own. That is a real marker of progress. But it reads less as a sudden break and more as the next rung on a ladder AI has been climbing steadily.
Continuity, not rupture
Three years ago, large language models struggled with basic arithmetic; only last year did they begin acing school-level competition maths. At January’s Joint Mathematics Meetings, researchers noted AI was contributing to genuine research, but only in tightly constrained settings that still required heavy human interpretation to reach a publishable result.
The Erdős proof fits that trajectory. The model drew widely on prior work to assemble a complete argument — playing to a clear AI strength, breadth of recall, rather than pioneering a new technique. Crucially, human mathematicians then cleaned up and extended the result, the same human-in-the-loop pattern seen across early AI-assisted science and, notably, in how the cyber-AI tools now reaching UK institutions still need expert checking of their output.
Looking forward
The likely medium-term shape is complementarity rather than replacement: AI models bring near-total recall of past work and limitless patience for tedious dead-end strategies, while humans still reason more deeply about a single problem and ask the better questions. For UK research institutions and businesses weighing AI’s role in knowledge work, the lesson is grounded rather than breathless — these systems extend expert capacity, they do not yet substitute for the judgement that turns raw output into something trustworthy.