OpenAI recently published hundreds of purported solutions to challenging mathematical problems, aiming to demonstrate the advanced capabilities of its AI models. However, an advisory group of prominent mathematicians known as the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), convened by Princeton’s Institute for Advanced Studies, has raised concerns that these proofs do not fully meet established standards in the field. Particularly, AGMAI emphasized the importance of human comprehension of mathematical results, which appears lacking in many of OpenAI’s released proofs, with less than half formally verified and only a fraction providing transparent reasoning steps.

The issue is further underscored by a new paper from researchers at the University of Cambridge and King’s College London. Their study highlights discrepancies between the natural language explanations generated by OpenAI’s models and the formalized proofs encoded in the Lean programming language, which is intended to verify mathematical correctness. This translation gap raises doubts about the reliability of AI autoformalization without human expert review, a critical step recommended by the advisory group to maintain rigor and trust in the mathematical community.

AGMAI also criticized OpenAI’s practice of testing proprietary models with open mathematical challenges, urging a halt to such experiments without open collaboration. They advocated for increased support for human mathematicians who can interpret, verify, and communicate AI-generated results, ensuring deeper understanding and integration into the broader scientific discourse. Mathematicians, including Fields Medalist Terence Tao, have pointed out that AI-generated proofs lack human insight, making it difficult for experts to engage with or build upon the findings effectively.

Ultimately, while OpenAI has committed to transparency by releasing many proofs and some data on reasoning processes, the consensus among experts is that more work is needed to bridge AI output and human verification. The mathematical community stresses that AI solutions should not only be technically accurate but also comprehensible and subjected to peer review principles that underpin scientific progress. This dialogue highlights ongoing challenges at the intersection of AI innovation and the highest standards of mathematical scholarship.