AI unsolved math problems gains trigger debate among mathematicians

AI unsolved math problems moved from benchmark talk to working research in 2026, after OpenAI published a counterexample to the Unit Distance Conjecture in May. The result disproved a geometric graph theory conjecture that had stood since 1946 and was among the open problems associated with Hungarian mathematician Paul Erdős.

The result was not isolated. Within a week, human researchers used the central proof idea to disprove another major conjecture. Since then, AI systems have been reported finding counterexamples, identifying patterns and helping convert arguments into machine-checkable proofs.

Epoch AI, which runs the FrontierMath benchmark, recently said it had recorded a second solution in its FrontierMath: Open Problems test, which draws on major unresolved questions. OpenAI also introduced its Astra model with ten mathematical solutions of varying difficulty, a launch that positioned the model squarely in automated reasoning.

Can AI solve unsolved math problems?

AI is now solving some open problems, especially in areas such as graph theory where search, counterexamples and formal verification can carry much of the work. It has not solved the hardest problems tracked by Epoch AI, and the remaining Millennium Prize Problems are still unsolved by both humans and AI.

That split is shaping the reaction inside mathematics. Abhishek Saha, a professor at Queen Mary University of London, wrote on X that frontier AI models in his research area are “at least as good as a solid and indefatigable PhD student.” He said one day with GPT-5.5 Pro handled routine work that previously would have taken weeks, leaving him more like a conductor than the full orchestra.

Trefor Bazett, writing in The Conversation, framed the current period as a possible new golden age for mathematics. He pointed to an April 2026 Carnegie Mellon University paper that solved an open Ramsey theory problem using a combination of SAT solvers, language-model-generated code and formal proof verification.

That optimism comes with a clock attached. Fields Medalist Timothy Gowers predicted in 2000 that computers could eventually handle enough proof work to transform pure mathematics. He now says GPT 5.6 Pro twice solved on its first attempt problems on which he had spent substantial time. Gowers wrote that the experience was unpleasant even though he welcomed the solutions.

Gowers’ concern is cultural rather than technical. He has warned that if mathematical literature expands faster than people can absorb it, the field could lose the human communities that understand and transmit deep expertise. He discussed those worries in connection with the Leiden Declaration on Artificial Intelligence and Mathematics, a standards effort backed by the International Mathematical Union and signed by more than 3,000 mathematicians. The declaration calls for transparency in AI use, respect for authors’ rights and human responsibility for results.

The limits are still visible. Bazett noted that AI performs unevenly across mathematical fields. Epoch AI’s hardest categories, “Major Advance” and “Breakthrough,” have no AI-solved problems so far. OpenAI’s Astra did not solve the six remaining Millennium Prize Problems, though OpenAI researcher Noam Brown has suggested more compute could change that.

Terence Tao, speaking at the 2026 International Congress of Mathematicians, described a future in which proof scarcity may turn into proof overload. In that scenario, mathematicians still have work to do: checking results, explaining them, deciding which ones matter and fitting them into broader theory.

Not every mathematician finds that reassuring. Kirwin Hampshire, writing on Substack, argued that treating AI as a tool misses what humans may lose if machines take over the creation of new mathematics. He described the Leiden Declaration as “a well-muffled scream” and called for more candid discussion about what the field is becoming.

This story draws on original reporting from The Decoder.