The Erdős precedent — what changes when AI-authored mathematics enters peer review
OpenAI proves a 1946 Erdős conjecture. Will Sawin refines the bound. The collaboration shape — AI produces, human reviewer refines — is the first concrete answer to 'who audits AI-generated mathematics.' That answer matters more than the specific theorem.
The result and the refinement
OpenAI's general-purpose reasoning model autonomously disproved Paul Erdős's 1946 planar unit-distance conjecture. The model produced both a new family of constructions beating the square-grid and a mathematical proof. Princeton's Will Sawin then refined the bound to δ ≥ 0.014.
The sequence — AI produces the construction and proof, human reviewer tightens it through standard peer-review-style refinement — is the methodological precedent. Mathematics has well-developed credentialing infrastructure for collaborative proofs; the AI-as-co-author shape slots into that infrastructure cleanly.
What it answers
The open question 'who audits AI-generated proofs?' has had vague answers for two years. The default response was some combination of 'the lab' (unsatisfying for community standards) and 'eventually, mathematicians' (true but unspecified). The OpenAI-Sawin sequence is the first concrete answer: any mathematician with subject-matter expertise reviews the proof; the review process is identical to human-authored work.
What this means for ICML, NeurIPS, and the major math journals
The 2026 submission cycle will likely see substantially more AI-authored research. Program committees need policies; the Erdős workflow provides a credibility template. Three implications:
- Authorship attribution. The OpenAI paper credits the model alongside human researchers. This pattern — explicit AI co-authorship rather than ghost-writing — will likely become the journal-level default.
- Reproducibility. AI-authored proofs need the same reproducibility standards as human-authored ones. The proof is the artifact; the AI is the tool that produced it. Standard peer-review treats both equally.
- Volume management. Journals may face order-of-magnitude submission volume increases. Triage processes will need adaptation; the math-community equivalent of code-review tooling will need to ship.
The under-noticed economic shift
Mathematics research has historically been a slow-moving field with stable credentialing and citation patterns. AI co-authorship at scale changes the economics — proofs become cheap to produce and refinement-to-publishable-quality becomes the bottleneck. Tenure committees, grant agencies, and prize committees all rely on slow-moving credentials; the Erdős precedent forces all of them to update.
The Erdős proof matters less because of what it solved than because it answered the question 'how does this work?' with a concrete workflow.
What to watch
The Q3 2026 NeurIPS submission cycle is the first major venue where AI-authored work will likely arrive at meaningful volume. The conference's policy response — and the community's reception of the response — will calibrate every other venue's stance. Mathematical journals (Annals of Math, JAMS, Inventiones) will follow with a lag.
For procurement teams not in the research-publication chain: AI models being formally credentialed as research co-authors elevates the brand-tier credibility of every lab that produces this kind of result. Procurement RFPs in 2027 will likely cite published mathematical results alongside benchmark numbers.
OpenAI — Erdős conjecture disproof → · TechBriefly — OpenAI Erdős proof → · TechCrunch — OpenAI 80-year-old math →