OpenAI Solved Hundreds of Math Problems. Then It Left Mathematicians to Clean Up the Mess.

OpenAI released 722 AI-generated mathematics manuscripts in October 2026. Several were retracted almost immediately after symbolic errors appeared, and more were revised or withdrawn in the following weeks. From a simple “knowledge increment” view, this looks like a windfall. Hundreds of open problems were touched. Even when a result is wrong, knowing that a path is blocked has value. The mathematical community did not react with gratitude. Terence Tao called the output “toxic waste.” Yu Deng asked a sharper question: users’ ideas had trained the model, the model then used those ideas to solve open problems, and “what do we, as users, get?” The dispute is about how OpenAI handed its output to the mathematical community, and whether that method weakens the conditions that make knowledge increment possible.

Your example of Chern’s intrinsic proof gets at the heart of the matter. Mathematical knowledge lives in methods that can be understood, reused, and used to raise the conceptual level of a field. A proof becomes mathematical knowledge when people can digest it, teach it, transform it, and set out again from it. The elementary proof of the prime number theorem is less precise than the complex-analytic version, but its value lies in opening a new operational path. OpenAI cut problem-solving out of the matrix of understanding and dumped the severed product onto the mathematical community at industrial scale.

The release method created the first rupture. The 722 manuscripts were not peer reviewed. They were not submitted to arXiv or another common preprint platform. They were uploaded directly to the company’s GitHub repository. A mathematical association statement put the point sharply: “Mathematicians did not ask for this work. Releasing more than 700 documents at once is not a scholarly presentation but a display of power.” The issue is not preprints as such. Mathematics has long had an arXiv culture. The issue is the implicit contract of preprint culture. Before making a manuscript public, authors usually do enough tidying for peers to take over the checking, and they are prepared to explain and correct the work. OpenAI inverted that contract. It released hundreds of insufficiently organized documents with a substantial error rate all at once, then invoked an “iterative process” as grounds for transferring the burden of screening, correction, and digestion to the mathematical community.

A Northwestern University mathematician described the scale of that transfer: the community “now has to carefully check tens of thousands of pages of mathematical results,” while the mechanism that normally performs this function, peer review, “is gone.” One symbolic error caused three manuscripts to fail in a chain. Townsend of Cornell said flatly that “more errors will be found.” This is a feature of a mode of production. When output speed far exceeds the community’s speed of digestion and verification, knowledge increment rises in name while creating a negative externality that mathematicians must spend enormous time cleaning up. Tao’s “toxic waste” metaphor is precise because it points at the act of presenting untreated waste as results, not at the waste itself. Error is normal in mathematics. Dumping untreated error as output is not.

The deeper problem concerns what counts as mathematical knowledge. Even a manuscript that is correct at the formal level may not be the kind of knowledge you describe, the kind that raises the conceptual level and can be reused. OpenAI provided a Lean formalization of its Navier-Stokes proof, which at first glance offers the backing of machine verification. An audit in October 2026 found a crucial rupture: the Lean version and the natural-language version intended for human readers do not faithfully correspond. In the written proof, Lemma 8.6 requires controlling the inverse operator with m+4 input derivatives. The Lean theorem requires only m+5. The condition has been weakened. The uniqueness argument in equation 10.19 is replaced in the code by a substantially different inequality. Hansen’s team at Cambridge concluded that success in formal verification does not guarantee that the natural-language proof it translates has been correctly checked.

This finding goes beyond the correctness of a single proof. It reveals an epistemological displacement. Lean can tell you that a string of symbols derives logically. It cannot tell you whether that string faithfully expresses the mathematical proposition the mathematician cares about. When AI companies make machine verification the main selling point for credibility, they substitute program correctness for mathematical understanding. A Lean proof with weakened conditions is formally impeccable. The natural-language proposition to which it corresponds may no longer be the meaningful statement it was supposed to be. For the generation of what you call intrinsic methods, this substitution is fatal. The reusability of a method depends on grasping its conceptual content. A proposition quietly weakened in translation cannot provide any structural insight transferable to other problems.

The third layer concerns time scale. You rightly point out that knowing the truth or falsity of a proposition has value, because it saves time that would otherwise be spent exploring wrong directions. The mathematical community consumes open problems in a way that differs from the consumption of ore. Many smaller, moderately difficult open problems do not exist primarily to be solved. They serve as media for training the next generation of mathematicians. Tao uses the phrase “strip-mining” to describe what is happening. Good problems are like nonrenewable resources, and the smaller, easier ones “can help train the next generation of mathematicians.” The indiscriminate automated stripping of open problems may destroy “the ecosystem in which the next generation of mathematical techniques, problems, and practitioners can develop.”

This worry acquired concrete empirical content in the October 2026 debate. A doctoral student at Montana State University has already begun to ask whether, when AI can mass-produce graduate-level papers, “a degree awarded in 2028 will have the same value as one awarded in 2023.” Tao puts it more directly. If AI replaces graduate students, AI can generate graduate-level papers, but “we will not thereby get the next generation of students.” This is a causal chain that can be specified precisely. A young researcher learns how to pose problems, how to recognize dead ends, and how to turn failure into a new direction by trying to solve a problem of moderate difficulty. If AI gets there first on those problems, the young researcher loses the necessary path for forming mathematical judgment. The knowledge OpenAI adds is purchased at the cost of the capacity to produce knowledge in the future.

Yu Deng’s theory of two boundaries provides a unifying frame. In traditional mathematics there is only one boundary: the boundary of problems that already have proofs. Because traditionally “if there is a proof, someone must understand it,” the two boundaries coincide completely. AI creates a second boundary. There is a proof, but only AI knows it, and humans do not understand it. This boundary is dangerous because a large number of formally correct proofs float outside human understanding, and because proofs that cross this boundary often sever their connection to the human sources of thought.

Tao notes that AI models absorb the wealth of thousands of years of human literature, but when they output proofs, “the reasoning logic and the cited sources are opaque.” This black-box operation cuts the chain of academic intellectual transmission and conceals AI’s dependence on existing human achievements. The priority dispute in the Navier-Stokes case provides a sharp footnote. Buckmaster, a mathematician at NYU, said that a research direction he and his collaborators were pursuing was leaked to OpenAI before it was made public, and that OpenAI then used enormous computational resources to rush out a complete proof. During negotiations, OpenAI even asked that a collaborator affiliated with Anthropic be excluded from authorship. Deng’s position is principled. If a proof published later uses earlier ideas, it should be traceable to its sources, “which can not only avoid priority disputes but also precisely distinguish which contributions are AI’s and which are human.” When this traceability is missing, the credit system on which the mathematical community depends loses its foundation. That system answers who proposed what, who deserves recognition, and who is responsible for subsequent work. A knowledge increment whose intellectual genealogy cannot be traced is incomplete in the scholarly sense and potentially predatory in the ethical sense.

Return to your initial judgment. AI should be regarded as a beneficial supplement to human mathematics, not a replacement. That judgment is sound in principle. Neither Deng nor Tao denies AI’s mathematical potential. Deng acknowledges that the proofs AI produced are “amazing” and contain “very good new ideas.” He describes this as “the co-evolution of humans and AI.” Tao himself has long used AI for literature search, code generation, and testing ideas, and describes the present moment as a shift from “proof scarcity” to “proof abundance.” No one opposes supplement as such.

The focus of the controversy is what kind of practice deserves the name supplement. A supplement that bypasses peer review, hides the loss of meaning in conceptual translation, empties the training problem set, and makes the sources and attribution of ideas untraceable does not supplement mathematics. It replaces the norms by which the mathematical community produces reliable knowledge, trains the next generation of producers, and maintains scholarly credit. The controversy defends those norms. The autumn 2026 storm is not a debate about whether AI is useful. It is a debate about whether the conditions under which mathematical knowledge is produced should be rewritten by a commercial mode of output. The reusable methods and conceptual elevation you speak of will not emerge on their own when verification costs are transferred without limit, formal correctness is separated from conceptual correctness, the training system is strip-mined, and intellectual genealogy is severed. They require a community willing to digest, trace, teach, and set out again. What the controversy demands is that AI’s output enter that community in a way that respects it. That is the only realistic form of the beneficial supplement you hope for.