AI Models Generate Novel Mathematical Proofs, Disproving Decades-Old Conjectures

For decades, artificial intelligence has treated mathematics as a closed-book exam. Models were trained to solve known problems, competing on olympiads and standardized tests where the answers were already established. As recently as 2025, the benchmark for AI was achieving gold-medal performance at the International Mathematical Olympiad—a formidable feat, but one that relied on curated problems with fixed targets.[1][2]

But in the spring of 2026, the paradigm shifted from retrieval to genuine discovery. AI is no longer just passing tests; it is actively expanding the frontier of human knowledge by generating novel proofs for open research questions that have stumped mathematicians for decades. This transition marks the moment artificial intelligence moved from being a sophisticated calculator to a genuine scientific collaborator.[6]

Two major breakthroughs have demonstrated that frontier AI models are now capable of producing original, research-level mathematical proofs. In February, Google DeepMind unveiled Aletheia, an agent powered by Gemini Deep Think, which autonomously solved open questions from the Erdős Conjectures database.[2]

Then, in May, OpenAI announced that its general-purpose reasoning model had disproved a central conjecture in discrete geometry tied to Paul Erdős's planar unit distance problem—a famously stubborn question first posed in 1946. The Erdős planar unit distance problem asks for the maximum number of pairs of points in a plane that can be exactly one unit distance apart.[1]

For nearly eighty years, mathematicians had relied on grid-based, geometric intuition to approach the problem, treating certain constructions as basically optimal. OpenAI's model upended this assumption entirely. Instead of relying purely on geometric grids, the AI drew from unexpected areas, including algebraic number theory, to construct a counterexample that beat the constructions mathematicians had long accepted.[1][4]

The mathematical community's reaction has been unprecedented. Fields Medalist Tim Gowers called the result a milestone in AI mathematics, positioning it as one of the first clear cases of an AI independently solving a famous open mathematical problem.[1]

Mathematician Daniel Litt noted that it is the most unique interesting result produced autonomously by AI so far, and the first AI proof that would likely be publishable in a top-tier math journal even if humans had produced it alone. The fact that the AI crossed disciplinary boundaries to find its solution is what makes the achievement so remarkable to veterans of the field.[1][3]

DeepMind's Aletheia achieved similar milestones in its own right. Unlike Olympiad problems, research-level mathematics requires navigating vast literature and avoiding the hallucinations that plague standard foundation models. To overcome this, Aletheia employs a natural language verifier that identifies flaws in candidate solutions, enabling an iterative process of generating, verifying, and revising.[2]

Crucially, the agent can admit failure, a feature that improves efficiency when collaborating with human researchers. Aletheia's capabilities extend far beyond pure combinatorics. It recently extended a theorem in economic theory—the Revelation Principle for auctioning AI generation tokens—by using advanced topology and order theory to accommodate continuous real numbers, invalidating the original proof that only worked for rational numbers.[2]

It also found novel analytical solutions to tricky integrals containing singularities, which are necessary for calculating gravitational radiation from cosmic strings in physics. These cross-disciplinary applications prove that the underlying reasoning engine is generalizing across the hard sciences.[2]

Despite these successes, the question of whether AI is truly creative remains a subject of intense debate among experts. The First Proof initiative, launched by the National Museum of Mathematics and featuring prominent mathematicians like Manjul Bhargava and Alex Kontorovich, aims to evaluate whether these systems are genuinely engaging in discovery or merely performing sophisticated combinatorial search.[5]

As Lauren Williams, a MacArthur Fellow and panelist for the initiative, noted, AI can sometimes produce brilliant solutions, but other times it generates answers that are nowhere near correct. The boundary between creative mathematical insight and brute-force computation is becoming increasingly blurred, forcing the community to redefine what it means to do mathematics.[5][6]

Furthermore, the originality of these AI-generated proofs is under intense scrutiny. If a model merely surfaces hidden but existing answers by connecting vast amounts of literature, its utility is immense but fundamentally different from human intuition. However, if AI models can assemble partially known ideas into something substantially new—as seen in the application of algebraic number theory to a geometric problem—they cross the threshold into true scientific collaboration.[1][4][6]

The implications for the future of mathematics are profound. AI is transitioning from a calculator to a co-author. As these models scale, they are expected to tackle even more abstract fields, such as algebraic geometry and topology, where human intuition often hits its limits. For now, the consensus is clear: AI has not solved math, nor has it replaced the human mathematician. But it has undeniably become a powerful scientific companion, capable of breaking century-old deadlocks and expanding the frontier of human knowledge.[2][5][6]