rationale:

Core capability vs. theory distinction

AI capabilities in mathematics have advanced astonishingly through 2026, culminating in the autonomous resolution of decades-old conjectures like forced Navier-Stokes blowup and Sogge's local smoothing conjecture. However, the resolution criteria demand an AI-introduced definition, structure, or theory that becomes an organizing concept—not merely the solution to a posed problem. Thus far, every accepted AI result operates within the "convex hull" of existing ideas What should we tell our students? - Terence Tao. An audit of 719 AI-generated manuscripts released by OpenAI in October 2026 found that only 0.3% introduced a notable new definition, compared to 5.3% in recent human arXiv math papers New definitions in the OpenAI math papers - David Holmes New definitions in the OpenAI math papers - David Holmes. Most AI definitions are engineered for a single theorem rather than offering broad theoretical utility.

The likely path to organizing concepts

The most plausible path to an AI-authored organizing concept is through reusable machinery built to attack existing problems. For instance, perfectoid spaces were originally introduced to attack the weight-monodromy conjecture The Perfectoid Concept: Test Case for an Absent Theory. AI is already demonstrating this pattern: OpenAI’s October 26, 2026 local smoothing proof relies on a "novel auxiliary PDE" and a conserved current https://openai.com/index/building-scientific-understanding-with-academic-communities/ OpenAI's local-smoothing proof claim: scope, Lean ..., which an expert described as "historic" Local smoothing conjecture. While that specific manuscript is ineligible due to the October 27 cutoff, it provides a clear template. Because humans are permitted to pose the problem and curate outputs, researchers and labs have a strong incentive to prompt frontier AI systems to develop and formalize these auxiliary structures into full-fledged frameworks.

Institutional digestion and adoption timelines

For an AI-generated concept to qualify, it must be adopted by multiple groups and feature in dedicated surveys, workshops, or courses. Human precedents suggest this adoption process can be swift when a concept is genuinely useful: perfectoid spaces had an arXiv survey and a dedicated workshop within roughly two years of introduction The Perfectoid Concept: Test Case for an Absent Theory [1303.5948] Perfectoid Spaces: A survey Perfectoid Spaces and their Applications. Deliberate institutional efforts are also accelerating the digestion of AI math. Initiatives like the $15M Simons/OpenAI/Anthropic program funding workshops https://www.simonsfoundation.org/news/simons-foundation-launches-15m-effort-to-help-mathematicians-navigate-ai-era/ and dedicated lab "shepherds" Working with academic communities to build scientific ... will systematically lower the barrier to community uptake. Crucially, the resolution date is the manuscript's initial release, meaning the 1.5–3 year social recognition lag affects when the question can be settled, but does not push the true occurrence date further into the future.

Key uncertainties and the right tail

The primary uncertainty is whether AI can bridge the gap from generating highly specialized concepts to producing elegant abstractions that simplify human researchers' thinking What's the Future for Pure Math Research in the Age of AI?. Organizing concepts are exceedingly rare—only a few emerge per decade across all of mathematics. There is also a distinct risk that human researchers will refine AI-generated concepts to the point where credit becomes blurred, or that community resistance to a flood of unread manuscripts buries genuinely useful definitions AHM Statement on OpenAI's October 6 Release of Mathematical ... https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/. Expected timelines for major open problems provide a useful comparison , but useful conceptual frameworks may follow a different trajectory from solutions to specific Millennium problems: reusable machinery could emerge before those solutions, while even major solutions need not yield an organizing concept. This supports a somewhat wider distribution rather than tightly coupling the two milestones. I put a 10% probability on a qualifying manuscript appearing by April 1, 2027, and 25% by November 1, 2027. These bottlenecks point to a median of September 1, 2028—aligning with expectations that AI could master theory-building by 2029 The End of Mathematics—with 75% probability by June 1, 2030, and 90% by January 1, 2034. The long right tail accounts for attribution failures, social resistance, or the possibility that conceptual invention remains stubbornly human.