At the intersection of AI and mathematics, a long-overdue redemption arc has finally concluded. On May 20, 2026, OpenAI announced that its general reasoning model had successfully proven the unit distance conjecture, posed by Paul Erdős in 1946, to be false — the first time in mathematical history an AI has independently completed a proof of an open problem and earned endorsement from domain experts.
Recap: from fiasco to comeback
Last October, then-OpenAI VP Kevin Weil loudly announced that GPT-5 had solved 10 previously-unsolved Erdős problems, only to be publicly rebuked by mathematician Thomas Bloom as a serious misrepresentation — the model had merely reproduced existing solutions from the literature, briefly costing OpenAI its standing in the academic world.
Seven months later, OpenAI struck back. This time, top mathematicians including Noga Alon, Melanie Wood, and Thomas Bloom co-published companion commentary validating the proof. Bloom had sharply criticized OpenAI's previous misstep, and his personal endorsement this time speaks volumes about the weight of the conclusion.
What does the technical breakthrough mean?
The model was not specifically designed for mathematics — it is a general reasoning model. It could navigate long chains of logic and cross-domain connections, discovering a brand-new construction superior to the traditional approximate square grid. As Bloom wrote: AI is helping us more completely explore the mathematical cathedral humans have built over centuries. What other undiscovered wonders are waiting?
Previous doubts about AI's math ability focused on hallucination and memorization. This time, hallucination was ruled out through independent mathematician verification, and the originality of the construction rules out memorization. This isn't the end, but it is a turning point — from solving math problems to doing math research.