OpenAI's math proofs fall short of field's standards

openai released 719 math solutions, claiming to have consulted an advisory group of elite mathematicians. however, the solutions fell short of the field's standards, particularly in human understanding of mathematical results. only 42% of proofs were formalized, and just 10 included the model's chain of thought. the advisory group had requested that openai stop testing advanced mathematical problems on proprietary models and include machine-readable metadata correlating natural language and formal artifacts. openai did not follow these recommendations, sparking concerns about responsibility and human involvement in math problem-solving.