The Great Unraveling: AI Reshapes Mathematics with Rapid Breakthroughs

The summer of 2026 marks a pivotal period for mathematics, as artificial intelligence models have rapidly disproven or advanced numerous open problems, far surpassing human progress over the past decade. What began a year ago with AI solving International Math Olympiad problems, initially dismissed as the “LeetCode of math,” quickly escalated. In May, an OpenAI model disproved the 80-year-old Erdős-Ulam conjecture on unit distances, signaling a new era. This was followed by Levent Alpe, assisted by Fable, disproving the 87-year-old Jacobian conjecture in algebraic geometry, a problem on Smale’s list of the hardest 21st-century challenges. Shortly after, the 30-year-old Dense Garments conjecture in graph theory fell to GPT 5.6, requiring only four prompts and a 58-word input to generate a counter-example. These rapid advancements have prompted widespread concern, notably from Fields Medalist Terence Tao, who, at the International Congress of Mathematicians, warned of a “crisis in the foundations of mathematical values and practices.” The Leiden Declaration, signed by 16 researchers from 15 universities, earlier called for guardrails around AI in mathematics research, a plea largely unheeded.

The pace of AI’s mathematical prowess has only intensified, with major tech labs now openly showcasing their capabilities. Two weeks ago, OpenAI revealed its next-generation model internally solved ten open problems across mathematics and theoretical computer science. These solutions, including an improvement to the general bound on sphere packing—a problem untouched since 1978—were formally verified with Lean certificates and published on GitHub. Not to be outdone, Anthropic announced last week that an unreleased model, prompted by BunJS creator Jared Sumar, made “groundbreaking” progress on the 167-year-old Riemann Hypothesis, one of the seven Millennium Prize problems. After generating 650 initial attempts, Claude, guided by Sumar, orchestrated 60 sub-agents, executing 2400 shell commands and writing hundreds of Python scripts, consuming 31 million tokens over a day and a half. While not a full proof, this effort bumped the fraction of solutions provably satisfying the hypothesis from 41% to 67%, a result validated by Anthropic mathematicians and external number theory experts, and also formalized in Lean. This unprecedented surge in AI-driven mathematical discovery signals a profound shift in how complex problems are approached and potentially solved.