AI cracked an Erdős math problem. Now experts want guardrails

An AI model developed by OpenAI has disproven a long-standing mathematical conjecture first proposed by Paul Erdős in 1946. The unit distance problem, which asks for the maximum number of pairs of points on a flat surface that can be separated by the same distance, was left unresolved for decades until OpenAI’s model generated a counterexample using tools from algebra and number theory. The breakthrough was published on May 20, with a detailed proof available on OpenAI’s website. While mathematicians like Melanie Matchett Wood of Harvard University hailed the discovery as a significant advancement for mathematics, they noted that the AI’s reasoning lacked the creative insight typically associated with major mathematical breakthroughs. The model’s approach relied on exhaustive exploration rather than innovative thinking, according to Wood and Thomas Bloom of the University of Manchester. OpenAI researchers, including Sébastien Bubeck, confirmed that the model was not specifically guided or trained for this task, yet it independently identified a flaw in Erdős’ conjecture. The proof was later reproduced by another researcher using a publicly available AI model, reinforcing concerns about the limitations of current AI capabilities in mathematical discovery. Despite the achievement, experts have raised alarms about the ethical and methodological risks of AI in mathematical research. A declaration signed by over 1,590 mathematicians, published on June 2, calls for stricter oversight to ensure responsible and verifiable use of AI in the field. The document warns that unchecked AI could undermine the rigor and transparency of mathematical work, threatening the integrity of scientific progress. The incident has sparked debate over AI’s role in advancing mathematics. While the technology demonstrates potential for solving complex problems, critics argue that its lack of creative intuition limits its ability to make groundbreaking discoveries. OpenAI acknowledges that AI currently excels at methodical problem-solving but struggles with the kind of innovative leaps that define major mathematical achievements.