AI Systems Solve Multiple Longstanding Mathematical Problems, Raising Questions About Future of Human Mathematics

In a striking series of developments in May 2026, artificial intelligence systems have solved several longstanding mathematical problems that have remained open for decades, prompting reflection on the future role of human mathematicians. An internal OpenAI model recently solved Paul Erdös's Unit Distance Problem from 1946, a central open problem in discrete geometry. Using algebraic number theory, the AI refuted Erdös's conjecture by constructing a set with n^(1+ε) unit-distance pairs, where ε is approximately 10^-38. Shortly after the AI breakthrough, human mathematician Will Sawin improved upon the construction to achieve approximately n^1.014 pairs, though the best-known upper bound remains n^(4/3), an improvement over Erdös's original n^(3/2).

The solution was notable for its efficiency: Scott Aaronson reports that the process appeared to be accomplished in a single attempt, with student Lijie Chen simply presenting the problem to GPT, which then produced a several-page mathematical argument that experts subsequently verified as correct. Shortly thereafter, DeepMind's AlphaProof Nexus system, which includes involvement from UT Austin colleague Swarat Chaudhuri, announced the settlement of nine additional Erdös problems, primarily in additive combinatorics, with proofs fully formalized in the Lean proof assistant. Most recently, an AI system called GPT5.5Pro solved a longstanding open problem in computer science concerning electrical flows on graphs.

These developments have prompted significant reflection within the mathematical community about the implications for human researchers. Aaronson notes that graduate students expressed concern about the future relevance of human mathematicians and scientists in a landscape where AI systems can solve fundamental problems. However, Aaronson cautions that selection bias may play a role—AI systems are likely failing to solve hundreds of other problems without announcement—and notes that human mathematicians had focused their efforts on proving Erdös correct rather than seeking counterexamples, a different research strategy. The AI's success may also reflect that the required expertise in algebraic number theory is uncommon among discrete geometers.

Despite these remarkable successes, uncertainty remains about the trajectory of AI mathematical capability. OpenAI accompanied its breakthrough with commentary from renowned mathematicians including Timothy Gowers and Noga Alon, reflecting on the AI's approach and methods visible through examination of its chain-of-thought reasoning. Aaronson expresses ambivalence about whether AI systems will soon solve major unsolved problems like P versus NP, noting that the role of human mathematicians could potentially reduce to selecting interesting questions and understanding AI-generated solutions, or whether systems will encounter fundamental limitations. He concludes that this question will be resolved empirically in the near term.