OpenAI數學論文評測:722篇齊發,3篇撤稿收場 | OpenAI Math Papers Review: 722 Dropped, 3 Retracted
By Kit 小克 | AI Tool Observer | 2026-10-19
🇹🇼 OpenAI數學論文評測:722篇齊發,3篇撤稿收場
722篇論文一次砍半公信力:OpenAI數學論文事件是什麼
2026年10月6日,OpenAI一口氣在GitHub的openai/math倉庫丟出722篇由內部未發布模型生成的數學手稿,宣稱涵蓋372個「結果家族」,平均每個結果耗費約3小時ChatGPT Pro運算。消息一出立刻登上Hacker News與數學圈社群熱議榜首,但隔天(10月7日)劇情急轉直下:OpenAI因為一個正負號錯誤(該是負一卻寫成正一),撤回其中3篇與「霍奇猜想」(Hodge conjecture,七大千禧年難題之一)相關的論文,這個錯誤還連帶拖垮另外2篇依賴該論證的手稿,最終722篇變成719篇,另有14篇被修訂、13處引用被更新。
42%通過Lean驗證?數字本身就有爭議
OpenAI宣稱42%的主要結果已用Lean定理證明器完成形式化驗證,但「形式化」這個詞在倉庫裡沒有明確定義,獨立社群重新計算後認為實際比例可能只有24%到36%之間。換句話說,超過一半的論文,驗證者只能靠人工讀證明、憑信任背書——而這正是AI數學推理現階段最大的罩門:模型可以產生看起來合理的證明,但人類審查跟不上產出速度。
陶哲軒開砲:問題不是對錯,是「沒人看得懂」
費爾茲獎得主陶哲軒(Terence Tao)在部落格公開批評這次發布,重點不只是放了3篇有錯的論文,而是整個發布流程完全沒有事先知會正在研究同一批難題的學界同儕,甚至打斷了幾位年輕研究者的既有專案進度。他更擔心的是結構性問題:當研究者只是對AI下指令、拿到解答,卻不夠理解證明本身時,數學論文的「作者」將逐漸失去向同行解釋、捍衛自己成果的能力。數學家Sophie Huiberts在陶哲軒部落格的客座文章中,更直接把部分論文評為「品質堪憂」、「一眼可拒」。
對工程師與一般AI用戶的啟示
這起事件不只是數學圈的茶壺風暴,而是所有依賴AI生成技術內容的人都該留意的警訊:
- 驗證永遠跑不過生成速度:AI產出的速度遠超過人類(或機器)審查的速度,量大不代表品質穩
- 「通過驗證」的定義要問清楚:連OpenAI自己的42%數字都經不起獨立複算
- 錯誤會連環爆:一個正負號錯誤拖垮3篇論文,AI生成的技術文件同樣可能藏著隱形的連鎖依賴
好不好用,試了才知道。
🇺🇸 OpenAI Math Papers Review: 722 Dropped, 3 Retracted
OpenAI Math Papers Controversy: 722 Proofs, 3 Retracted in a Day
On October 6, 2026, OpenAI dumped 722 AI-generated mathematical manuscripts onto its public GitHub repo openai/math, produced by an unreleased internal model across 372 "result families," averaging roughly three hours of ChatGPT Pro compute per result. It immediately topped Hacker News and math-community discussion — then, the very next day, things unraveled: OpenAI retracted 3 papers tied to the Hodge conjecture (one of the seven Millennium Prize Problems) after discovering a sign error — a +1 that should have been -1 — that invalidated a core stabilization-trace argument. The error cascaded into 2 more dependent manuscripts, shrinking the total from 722 to 719, with 14 papers revised and 13 citations updated.
Only 42% "Verified" — and Even That Number Is Disputed
OpenAI claims 42% of its headline results are formalized in the Lean theorem prover, but "formalized" is never precisely defined in the repo, and independent recounts put the real figure closer to 24-36%. That means well over half the papers rely on human reviewers reading and trusting the proof by eye — which is exactly where AI math reasoning is weakest right now: models can produce plausible-looking proofs faster than humans can check them.
Terence Tao Real Complaint: Nobody Understands the Proofs Anymore
Fields Medalist Terence Tao publicly criticized the release — not just for the 3 flawed papers, but because OpenAI gave zero advance notice to researchers already working on the same open problems, abruptly overtaking some junior researchers ongoing projects. His deeper worry is structural: when a researcher just prompts an AI and receives an answer without fully understanding the proof, the "author" loses the ability to defend the result to skeptical peers. Mathematician Sophie Huiberts, in a guest post on Tao blog, went further, calling some of the proofs "easy rejects" on quality grounds.
What This Means If You Use AI for Technical Work
This is not just inside baseball for mathematicians — it is a warning for anyone trusting AI-generated technical output:
- Verification cannot keep pace with generation: AI produces content far faster than humans (or even automated checkers) can validate it
- Ask what "verified" actually means: even OpenAI own 42% figure did not survive independent recounting
- Errors cascade: one sign error took down 3 papers — AI-generated technical docs can hide the same kind of chained dependency
好不好用,試了才知道。(You only know if it actually works once you have tried it.)
Sources / 資料來源
- Fortune: OpenAI math controversy over 370+ solutions
- Tom's Hardware: AI causes uproar in mathematics
- KuCoin News: OpenAI retracts 3 AI math papers due to sign error
延伸閱讀 / Related Articles
- OpenCode資安事故評測:API金鑰外洩,1.4%用戶遭盜刷 | OpenCode Breach Review: API Keys Leaked, 1.4% Users Hit
- Replit預設模型評測:棄用Claude,改押OpenAI | Replit Default Model Review: Drops Claude for OpenAI
- VS Code Copilot評測:Commit被偷加AI共同作者,微軟急回滾 | VS Code Copilot Review: Silent Co-Author Commits, MS Reverts
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言