Key Takeaways
- OpenAI falsely claimed that GPT-5 had solved 10 unsolved Erdős problems and made progress on 11 others.
- Mathematician Thomas Bloom clarified that the problems weren’t unsolved — GPT-5 had simply located existing solutions.
- OpenAI executives Kevin Weil and Sebastien Bubeck deleted or walked back their posts after widespread backlash.
- Industry leaders Demis Hassabis (DeepMind) and Yann LeCun (Meta) sharply criticized the incident.
- The controversy reignited concerns about AI hype, scientific verification, and credibility in high-stakes research claims.
OpenAI has come under fire after prematurely declaring that its GPT-5 model had achieved a “historic mathematical breakthrough.” The controversy erupted when Kevin Weil, OpenAI’s Vice President, announced on social media that GPT-5 had “found solutions to 10 previously unsolved Erdős problems and made progress on 11 others.” His post quickly went viral, amplified by OpenAI researcher Sebastien Bubeck, who triumphantly proclaimed that “science acceleration via AI has officially begun.”
The celebrations, however, were short-lived. Within hours, mathematician Thomas Bloom, curator of ErdősProblems.com, debunked the claims, explaining that GPT-5 had not solved any new problems at all. Instead, it had identified existing published solutions to problems that Bloom had personally marked as “open” on his website. “GPT-5 found references I was unaware of,” Bloom clarified, adding that his site’s “open” label did not necessarily mean the mathematical community considered the problems unsolved.
The revelation triggered widespread backlash from both the academic and AI communities. Demis Hassabis, CEO of Google DeepMind and a Nobel Laureate, dismissed the episode as “embarrassing,” while Meta’s Yann LeCun delivered a scathing critique, quipping that OpenAI had been “hoisted by their own GPTards”—a wordplay on the phrase “hoisted by your own petard.”
Bubeck later conceded the error, admitting that GPT-5 had merely retrieved literature references rather than generating novel proofs. His defense—that “literature search itself is a challenging reasoning task”—failed to quell criticism. Many observers argued that confusing database retrieval with genuine discovery risks undermining public trust in AI-driven science.
The controversy underscores a deeper issue in the race between AI giants like OpenAI and DeepMind to prove their systems’ reasoning capabilities. Both companies have achieved legitimate milestones this year, with AI models earning top scores at the International Mathematical Olympiad. Yet, incidents like this highlight the dangers of overhyping incremental progress and blurring the line between true scientific discovery and computational assistance.
As one industry analyst put it: “If AI companies start mistaking literature searches for new knowledge, credibility will become the next unsolved problem.”
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →