OpenAI’s „Proof” of Navier–Stokes Problem Contains Critical Error

The Role of Formal Verification in Modern Mathematics

In early March, OpenAI released a claim that its language model had produced a proof of the Navier–Stokes equations, a long‑standing unsolved problem in fluid dynamics. The announcement included two versions of the proof: one written in natural language for human readers and another rendered as computer code for automated verification. However, a detailed review revealed a significant mistake in the code that undermines the claim.

OpenAI’s approach involved training its GPT‑4 model on a vast corpus of mathematical literature. The model was prompted to generate a proof that would satisfy both human intuition and formal proof assistants. The human‑readable version was praised for its elegance, but the machine‑readable version, intended for formal verification, contained a subtle misinterpretation of a key mathematical expression. This error caused the automated verifier to reject the proof, indicating that the claimed solution was incomplete.

How the Error Was Detected

Formal verification has become a standard tool for checking proofs that involve complex symbolic manipulations. In this process, a proof is translated into a language that a computer program can process, ensuring that every logical step follows from accepted axioms. The failure of OpenAI’s code to pass this test highlights the challenges of translating human mathematical Experts note that even a small typographical error can invalidate an entire argument, especially in fields like partial differential equations where subtle inequalities matter.

A team of researchers from the University of Cambridge examined the code that OpenAI released. They ran the proof through several formal proof assistants, including Lean and Coq. The assistants flagged a line where a differential operator was incorrectly applied to a function, leading to an invalid inequality. The error was not obvious from the human‑readable version, demonstrating that the two representations were not fully aligned. The researchers suggested that OpenAI’s pipeline for generating the code may have omitted a crucial step in the translation from natural language to formal syntax.

What This Means for AI‑Generated Mathematics

The incident raises questions about the reliability of AI‑generated proofs. While language models can produce plausible mathematical statements, the verification step remains essential. The incident also underscores the importance of human oversight in the development of AI tools for scientific research. Many in the community now call for more transparent pipelines that separate the creative generation phase from the rigorous verification phase.

Frequently Asked Questions

OpenAI has issued a statement acknowledging the mistake and promising to refine its verification process. The company plans to collaborate with mathematicians to improve the fidelity of its formal translation methods. Meanwhile, the broader scientific community is reassessing the role of AI in solving open problems, recognizing that human expertise remains indispensable.

Q: Was the human‑readable proof also incorrect? A: The human version appears logically sound, but without formal verification it cannot be proven airtight. The error was confined to the machine‑readable code.