The GPT-5 Controversy Over Claimed Solutions to Erdős Math Problems
A critical analysis of OpenAI's claims about GPT-5 and the Erdős problems, focusing on AI's real role in mathematics.
Criticism and Clarifications on the GPT-5 Claim and the Erdős Problems
OpenAI recently promoted alleged successes of GPT-5 in solving complex math problems, in particular the well-known Erdős problems. However, many authoritative voices in the fields of artificial intelligence and mathematics have criticized these claims as misleading and inaccurate.
The True Nature of GPT-5's "Solutions"
Initially, an OpenAI executive stated that GPT-5 had found solutions to numerous Erdős problems, some of which had been considered open for decades. The statement generated excitement but was later largely walked back. A mathematician explained that GPT-5 did not actually solve the problems; instead, it located references to existing solutions in the vast body of literature online, which the researcher had not found until then.
Reactions and Official Retractions
OpenAI's claims were promptly criticized by other industry leaders, including the CEO of DeepMind, who called the situation "embarrassing". In addition, the owner of a website dedicated to Erdős problems clarified that the term "open problem" simply means that no documented solution is known to the author, not that the problem is unsolved. This led to an important conclusion: GPT-5 did not discover new solutions, but only retrieved references that already existed online.
The Real Value of GPT-5 in Mathematics
While these statements do not indicate an innovative resolution of the problems, they do highlight GPT-5's role as a research support tool, especially in finding scientific literature that is scattered or hard to access. According to some mathematicians, the use of AI can speed up text review and information aggregation, contributing to faster progress even if it is not directly problem-solving.
Conclusions and Key Points
- OpenAI's claims about solving Erdős problems with GPT-5 were retracted and judged misleading.
- GPT-5 identified existing references rather than making innovative discoveries.
- AI's role in mathematics is most useful as a support and accelerator for research activities, rather than as a solver of unsolved problems.
Going forward, it is important to communicate clearly the real limits and potential of AI technologies, to avoid misunderstandings and maintain scientific credibility.