Tao Zhexuan questions OpenAI: 719 AI mathematical proofs that humans really understand?陶哲轩质疑 OpenAI 公司：719 篇 AI 数学证明，人类真的理解了吗？

In October, OpenAI released 719 AI mathematical solutions, covering 372 mathematical result families, involving hundreds of open research questions, but after the release, it was accused of not fully meeting the standards proposed by the "Mathematical and Artificial Intelligence Advisory Group" (AGMAI). IT House previously reported that OpenAI initially published a total of 722 AI-generated mathematical manuscripts on GitHub on October 6, involving 372 outcome families and multiple open research questions. However, on October 7, 3 copies were withdrawn due to a symbol error and its chain effect, and the number of public catalogues is now 719. OpenAI said it consulted with the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), chaired by the Princeton Institute for Advanced Study, and consulted its public recommendations prior to the launch. However, from the public information, this release has not fully met the standards proposed by the group. AGMAI previously recommended that cutting-edge AI laboratories stop testing difficult mathematical problems on proprietary, inaccessible models and require disclosure of model names, prompts, inference chains, time-consuming, and computational costs. However, OpenAI still uses proprietary models, and only 10 manuscripts have model inference chains. In terms of human comprehensibility, the panel emphasized that AI-generated proofs should be easy for mathematicians to review and learn from; however, about 42% of the proofs published by OpenAI were reportedly not formalized and did not provide machine-readable metadata linking natural language proofs to formalized products. AGMAI also states that its advisory role does not constitute an endorsement of OpenAI's acquisition or publication of these results and that ultimately it is still up to the mathematical community to assess whether the recommendations have been fully implemented. This incident once again triggered controversy over the transparency, formal verification, peer review, and autonomy of academic research of AI-generated mathematical results. Tao Zhexuan posted on mathstodon on October 7, saying that he is not opposed to AI doing mathematics and has even been an active user of AI-assisted research; but he strongly opposes "using AI to quickly solve famous problems" as the main goal or product display. He believes that this will undermine the understanding, teaching, cooperation and open exploration mechanisms on which the mathematical community operates. Tao Zhexuan pointed out that in traditional mathematics, the breakthrough of a long-term conjecture is not just to “get the answer”: the author will make a report, participate in a seminar, communicate with his peers, and then prove to be simplified, interpreted, included in the textbook, and spawn follow-up questions, collaborators, and new research directions. The current practice of companies such as OpenAI is often driven by prompters to solve problems autonomously; once the goal is "solved", subsequent understanding, reporting, peer discussion and domain construction rarely occur. Tao believes that the publisher may not even be enough to explain the AI's output, answer questions, or interact with the field. He called this situation the “proof of indigestion” in mathematics: AI can generate propositions, proofs, and counterexamples at high speeds, but humans are too late to verify, understand, write, teach, and absorb, resulting in a large number of “proofs that no one can digest.” He is particularly concerned that OpenAI will use millennial puzzles such as the Navier-Stokes equation as benchmarks for model capabilities. In his view, this "how many problems to solve, how fast to solve" as a measure of "understanding and insight"; but speed and quantity itself is not equal to mathematical understanding. Tao Zhexuan stressed that once the problem is publicly "solved", it is almost impossible to return to an "unsolved" state. Even if people later want to find different paths and refine new methods from them, just knowing that the answer exists will "pollute" the exploration process. Therefore, he criticized this as an unsustainable large-scale “harvest”: treating open problems as a resource that can be consumed in bulk will eventually make the entire field of mathematics less fertile than under traditional research methods. IT House Note: Tao Zhexuan is one of the most prestigious mathematicians of our time, a professor of mathematics at the University of California, Los Angeles (UCLA), and a 2006 Fields Medal winner; often referred to as the “Mozart of mathematics.” Related Reading: OpenAI Releases Another Batch of AI Mathematics Research Results, Solving Hundreds of Unresolved Problems