OpenAI retracts 3 and revises 14 of 722 AI math manuscripts a day after release, raising reliability concerns (report)

What happened

ZDNet Korea reported this, citing TechCrunch. On October 6, OpenAI had an unreleased internal model work on about 4,000 math problems and published the results as 722 manuscripts. A day after release, a sign error in a calculation was found in one manuscript. That manuscript and two others built on it, three in total, were retracted. OpenAI also revised 14 manuscripts to strengthen proofs or correct explanations. The research areas include algebra and mathematical logic. Some mathematicians said the AI-generated solutions are hard to verify and questioned whether the work was released without enough review.

Provider claims

OpenAI said the AI found answers to problems mathematicians had long been unable to solve, and also wrote solutions explaining why those answers are correct.

Why it matters

Earlier reports focused on objections from mathematicians and alleged guideline violations. Now there are concrete numbers on manuscripts retracted and revised because of actual errors. The central question has become who verifies research results that AI produces at scale, and how.

Sources

• Press 오픈AI, 수학 난제 해결 주장…학계, 신뢰성에 의문 드러내 ZDNet Korea

Story thread

1. 2026-10-07 OpenAI publishes results on open math problems from an internal frontier model, with Lean-formalized proofs
2. 2026-10-08 Terence Tao shares 'Association for Human Mathematics' statement urging mathematicians to stop working with OpenAI (community report)
3. 2026-10-08 Terence Tao on OpenAI's math results: piling up answers won't keep pace with research community growth; Lean-verified improvements also shared (community report)
4. 2026-10-09 OpenAI retracts 3 and revises 14 of 722 AI math manuscripts a day after release, raising reliability concerns (report)
5. 2026-10-09 OpenAI's flood of math proofs departed from guidelines set by its own advisory mathematicians, report says

More about OpenAI

• 2026-10-09 P1 Other Unconfirmed Language models OpenAIAnthropic
  OpenAI tells investors its ARR neared $50 billion at end of September, below the reported $68–70 billion; big tech stocks swing (report)

  Korean outlet AI Times, citing the Financial Times (FT), reported on October 8 (local time) that OpenAI disclosed at a recent investor...

• 2026-10-09 P1 Policy Unconfirmed Safety & alignment OpenAI
  Three fired OpenAI safety researchers dispute misconduct allegations in open letter, warn of 'chilling effect' on safety culture (report)

  Three safety researchers fired from OpenAI last week published an open letter. They disputed allegations that they mishandled sensitive...

• 2026-10-09 P2 Research Unconfirmed Science & bioLanguage models OpenAI
  OpenAI's flood of math proofs departed from guidelines set by its own advisory mathematicians, report says

  TechCrunch reported that the proofs OpenAI released in bulk departed from guidelines set by a group of mathematics researchers OpenAI...

• 2026-10-09 P3 Product Generally available AI platforms & solutionsAgents OpenAIOracle
  OpenAI shares how Oracle uses ChatGPT Work and Codex in recruiting, engineering, and operations

  OpenAI showcased how Oracle used ChatGPT Work and Codex to turn expert knowledge into fast, repeatable workflows in recruiting,...

• 2026-10-08 P2 Other Unconfirmed Language modelsData & evaluation SemiAnalysisAnthropicOpenAI
  SemiAnalysis: on the same $200/month plan, Claude Opus 5.5 usage value is over 5x that of GPT-6.1 Sol (report)

  According to AI Times, SemiAnalysis on the 6th (local time) published its own measurements of the actual usage limits of major AI...