The Verification Problem (On OpenAI's Erdős Disproof) (korbonits.com)

OpenAI recently announced that its AI model successfully disproved Paul Erdős's planar unit-distance conjecture, a significant achievement in the AI and mathematics community. The model produced a proof that not only refuted the conjecture, but did so by generating an infinite family of point configurations that yield a polynomial improvement over previously established limits. This proof was rigorously verified by a team of nine mathematicians, including Thomas Bloom, who had previously criticized OpenAI's earlier claims of solving Erdős problems. The transition from an unverifiable claim to a verified proof underscores the importance of trust and human verification in AI-generated outputs. The new result highlights a crucial gap in the AI landscape: while AI can generate convincing mathematical proofs, it does not inherently possess the capability to ensure their correctness. The verification process, which involved multiple layers of human expertise, was the truly resource-intensive step. As AI's ability to produce plausible but incorrect proofs increases, the challenge of verifying these outputs becomes paramount. The article warns that reliance on human verification may not scale, suggesting a pressing need for formal verification systems that can ensure correctness independently of human oversight. This scenario advocates for a future where automated verification becomes essential, transforming the landscape of mathematical research and AI’s role within it.