OpenAI Proved the Quasi-Riemann Hypothesis

Summary

OpenAI's AI system produced a proof for a semi-Riemann hypothesis that narrows the strip where the zeros of the zeta function can be found. Human number theorists describe this as the most significant AI-mathematics discovery to date and are currently reviewing the proof. Two versions exist: a 199-page, Lean-verified proof (placing zeros in the [1/8, 7/8] interval) and a 49-page, unverified proof (placing zeros in the [1/12, 11/12] interval). The community is currently focused on the latter.

Why it matters

This marks the first time an AI has produced a result in a field requiring conceptual depth, such as analytic number theory. Rather than inventing new objects, the proof combines existing techniques, suggesting that the AI is capable of genuine mathematical reasoning rather than mere pattern matching. According to Radziwiłł from NYU, the semi-RH result has 'surprising' implications for the distribution of prime numbers. Furthermore, the fact that this work builds on the research of Radziwiłł and Dunn blurs the line between human and machine discovery. This also sets a new benchmark: if AI can solve a weaker version of a 160-year-old problem, the full Riemann hypothesis may no longer be out of reach for the same tools.

Highlights

• The proof is described as a 'surprising synthesis of known ideas' (Dunn, Georgia Tech); it is not a new conceptual breakthrough.
• Both papers are considered 'poorly written' by journal standards (Radziwiłł); the 199-page verified version and the 49-page unverified version tell different stories regarding rigor.
• Mathematicians are prioritizing the weaker 1/12–11/12 result over the stronger 1/8–7/8 result. This raises the question of which claim is actually reliable.

Critical take

The article relies on the superlative claims of a small group of mathematicians, despite the fact that the second (and currently preferred) proof is unverified; this makes the 'surprising' framing somewhat premature. There is also an uncomfortable asymmetry: the stronger result is 199 pages long and formally checked, while the weaker one is 49 pages and unchecked. Yet the community is betting on the weaker one. The article’s celebratory tone overlooks this detail.