OpenAI vs Mathematicians: Inside the Escalating Feud Over AI and Intellectual Work
TL;DR
- Twenty-five prominent mathematicians signed an open letter in early September 2026 accusing OpenAI and other AI labs of scraping proofs, preprints, and problem sets without consent, credit, or compensation.
- The feud is escalating as new AI models now solve Olympiad-level and research-level math, fueling fears over academic credit, peer review integrity, and the devaluation of human mathematical labor.
- The fight could reshape AI training rules and math publishing, with calls for opt-out datasets, proof provenance watermarks, and new licensing models for technical research.
The Letter That Shook the Math World
The math community has had enough. On September 4, 2026, twenty-five leading mathematicians published an open letter accusing major AI labs, with OpenAI named most directly, of building powerful reasoning systems on top of their intellectual work without permission.
Posted on a dedicated site and amplified across arXiv, MathOverflow, and X, the letter claims frontier models were trained on hundreds of thousands of papers, lecture notes, textbooks, and forum solutions scraped from the open web. The signatories, who include Fields Medalists, journal editors, and professors from Princeton, MIT, Cambridge, Oxford, and the IHES, argue this is not fair use. It is, in their words, industrial-scale appropriation of decades of publicly shared but not freely exploitable scholarship.
The core complaint is simple: mathematicians spent careers posting preprints for other humans to read, not for corporations valued in the hundreds of billions to ingest as training fuel.
What They Are Actually Accusing AI Labs Of
The letter breaks its case into three main arguments.
First is consent and compensation. Unlike novels or music, most math papers were posted to arXiv under licenses that allow sharing, not commercial AI training. The mathematicians say labs ignored robots.txt, paywalls for journals, and Creative Commons non-commercial clauses to build math-heavy corpora.
Second is credit and erasure. Modern AI models can reproduce entire proofs, invent near-identical lemmas, and solve homework and competition problems in seconds, often with no citation to the original authors. For a field where a single theorem can define a career, the signatories say AI-generated mathematics without provenance destroys the academic credit economy.
Third is contamination of the research ecosystem. The letter warns of a flood of plausible-looking but subtly flawed AI proofs overwhelming peer review, referees, and sites like arXiv and StackExchange. Several signatories cite personal experiences of finding their own unpublished work or referee reports paraphrased in chatbot outputs.
Why This Feud Is Exploding Right Now
Tensions have simmered for years, but three recent developments pushed mathematicians to go public.
The first was capability shock. In July 2026, OpenAI and Google DeepMind both announced models achieving gold-medal performance on the International Mathematical Olympiad, followed quickly by strong results on FrontierMath and new autonomous research benchmarks. What was once a party trick is now approaching research-assistant level, capable of suggesting lemmas for unsolved problems in combinatorics, number theory, and topology.
The second was the data debate. Investigations this summer alleged that AI labs used large-scale crawls of arXiv LaTeX source, Math StackExchange threads, and pirated textbook PDFs to train reasoning models. OpenAI has said it respects opt-outs and uses licensed and publicly available data, but has not released a full dataset list for its latest models.
The third was money. With OpenAI launching a dedicated Math Tutor and research copilot tied to its enterprise and education push, mathematicians say they are watching their free labor get repackaged as a $20-a-month subscription product.
OpenAI's Response So Far
OpenAI has not directly responded to each claim, but in a brief statement on September 8, the company said it takes creator concerns seriously, supports the development of provenance standards, and is continuing to expand publisher partnerships and opt-out tools for websites.
That has done little to calm critics. Similar statements in the past over news articles, books, and art did not stop major copyright lawsuits, several of which are still moving through U.S. and EU courts in 2026. Mathematicians signing the letter say they are now consulting with academic societies and legal scholars about next steps, including a potential collective licensing demand via the American Mathematical Society and European Mathematical Society.
It Is Bigger Than OpenAI
While the headline names OpenAI, the letter explicitly calls out the entire frontier lab ecosystem, including Anthropic, Google DeepMind, Meta, xAI, and DeepSeek. All have raced to showcase math reasoning as proof of general intelligence.
Many researchers privately admit they use these same tools daily to brainstorm proofs, debug LaTeX, and explore counterexamples. That uncomfortable duality, dependent on AI while feeling exploited by it, is why the debate is so bitter. As one signatory wrote, We love what these machines can do. We object to how they were built.
What This Means for the Future of Math Research
The immediate fallout could hit where math actually lives: arXiv.
Some prominent mathematicians are already proposing a delayed-publication model or a no-train registry where authors can flag preprints as off-limits for AI crawling. Others want arXiv to adopt machine-readable licensing that AI companies must legally respect, plus cryptographic proof watermarks to distinguish human theorems from AI-generated conjectures.
University departments are also scrambling. Princeton, Cambridge, and several others have issued new guidance this month requiring disclosure of AI assistance in submitted papers and theses, and some journals are piloting AI-proof detectors and stricter requirements for full human-verified proofs.
Longer term, the letter demands a new economic model: either pay for high-quality technical corpora through collective licensing, similar to deals struck with news publishers, or build open, mathematician-governed training sets where contributors are credited and compensated.
The High Stakes for AI Training
For AI labs, math is not just another data category. It is the holy grail for reasoning. The ability to prove theorems reliably is seen as the clearest path to trustworthy scientific AI, automated coding, and engineering breakthroughs.
If mathematicians successfully wall off fresh preprints, problem sets, and peer review data, labs could face a data bottleneck just as scaling laws for reasoning are taking off. That is why insiders expect a compromise attempt: paid licensing pools, citation engines that link AI proofs back to source papers, and researcher access programs offering free compute in exchange for data.
Whether that will satisfy a community built on open sharing remains unclear. For now, the open letter ends with a warning that has reverberated far beyond mathematics: If proving theorems becomes just more content to be scraped, the next generation may stop sharing them at all.
Get All The Latest Updates Delivered Straight To Your Inbox For Free!