一次性发布 700 多个文件并非学术成就的体现,而是权力的炫耀。
笔者敦促数学家们停止与 OpenAI 的合作,回归以人类理解为中心的科学理念。
这是人类数学协会在 OpenAI 发布大量数学研究成果后,专门写了一篇声明用来回应这件事,协会在声明中评论 OpenAI 做这件事是「数学家们并没有要求开展这项工作」,并呼吁所有数学家从现在开始都抵制 OpenAI。

声明一发,AI 和数学界的分歧又在 X 上争论得沸沸扬扬。
华人数学家,菲尔兹奖得主陶哲轩在自己的个人博客转载了这篇声明。很快就被网友转载为这份声明是由陶哲轩领衔发出,但事实是就连陶哲轩本人,都没有浮现在这个人类数学协会的会员名单里。
于是有人开始深挖这个人类数学协会是什么来头,网友们指责该协会和 OpenAI 同样不负责,尤其是所谓的「数学家们并没有要求开展这项工作」言论,被认为「是我最近读过的最滑稽、最荒谬的说法。」
笔者也发现比这份声明更有意思的是,人类数学协会的会员承诺。

要成为人类数学协会的会员,你必须做出三项承诺:
- 成员承诺不向商业人工智能公司给出工艺劳动力、数学知识、咨询或宣传服务。
- 成员承诺不发表由人工智能模型生成的数学文本(涵盖论文、评论、审稿报告、讲义、教学材料和书籍)。
- 反对人工智能小组的成员在研究工作中不使用人工智能模型。
总结下来就是,你要加入人类数学协会,就请你放弃 AI。
但更滑稽的是,有人把这份声明丢给 AI 检测工具 Pangram,发现检测结果是 100% 由 AI 生成。

另一边对 OpenAI 的抨击也没停止,所谓的 722 篇论文,只有 162 篇论文的主要结果经过了计算机验证。
发布第二天,OpenAI 还撤下了 3 篇手稿:一处符号错误使一个核心论证失效,并连带影响两篇依赖它的论文;另外 14 篇被修订,13 篇更新了引用关系。

OpenAI 自身也承认并非所有论文都使用了 Lean 形式化方法,在他们公开的 GitHub 仓库 README 文档里提到「一些未形式化的结果可能存在难题。笔者将努力尽快修复任何此类难题。」
也就是说,他们发表的很多内容可能都是错误的。

所以,到底应该听谁的,是跟着数学界一起呼吁所有的数学家都不要再使用 OpenAI 了,还是和 OpenAI 一起,接受 AI 带来的变化,接受 AI 即将颇为钟就能造出一个菲尔兹奖的现状。
一场抵制 OpenAI 的数学家联盟
这份完整的声明里,除了提到「停止与 OpenAI 的合作」和「数学家们没有要求 OpenAI 开展这项工作」外,也借这个事件讨论了数学研究的基本规则,即作者是否理解证明、是否亲自核验、谁承担责任、论文如何引用和同行评审。

这也是反对 OpenAI 声音里,讨论最多的一项争议。
在批评者看来,OpenAI 并不理解他们所发布的东西,甚至他们自己都看不懂那些 AI 撰写的论文和报告,只是看到 AI 得出了某个结论,就把它公诸于众,然后说自己应对了哪些数学难题。

另一维度,声明里还提到了 OpenAI 权力的难题,一个公司用公众无法访问的内部模型批量产生研究成果,再决定什么时候发布、发布什么,是否会让公司取得设定数学研究议程的权力。
有网友说,这份声明很可能会被列入将来讲授「AI 进入知识生产的早期历史」大学课程的一手材料。
声明背后,这个协会的信息也被网友扒出来:成立于 2026 年 8 月,当下仅有 809 名会员,且大多数都只是研究生,并非那些有过关键数学研究成果的数学家们。

809 名成员中,有将近三分之一是法国人,另外三分之一是美国人 ,其中还有 11 人来自中国。
类似的人类数学协会就像早些年 AI 刚刚出来的 AI 福利协会、AI 安全协会、AI 人类协会之类,除了这个 809 人的人类数学协会,还有由 9 位资深数学家组成的数学与人工智能咨询小组,以及由 28 位菲尔兹奖得主签署的数学宣言等。
mathandai.org – 由 28 位菲尔兹奖得主签署并得到 8000 多人兼容的《人工智能在数学领域严重失衡》宣言。
agmai.org – 数学与人工智能咨询小组,由高等研究院 (IAS) 主办:9 位资深数学家,独立且无偿,主要就实验室应如何发布人工智能生成的结果发表评论和建议。
ahmath.org – 人类数学协会,约有 800 名成员。所有成员都承诺不与商业人工智能公司合作,也不发表人工智能生成的数学文本;其中大多数成员(非人工智能小组)也拒绝在研究中使用人工智能。

上下滑动查看更多内容
这些数学家的观点不尽相同,但大多数都认为眼下数学确实受到了 OpenAI 的冲击,数学界面临着新的迷茫和困惑。
在社交媒体上,有人说 OpenAI 公布这些论文手稿,完全只是为了自身经济利益,而全然不顾数学界现有的结构,即署名论文作者、向期刊投稿以及经过同行评审等数学规范流程。
如果 OpenAI 想要摧毁数学界,这无疑是一个绝佳的途径。
未应对的难题是数学界历经数十年甚至数百年积累的宝贵资源。它们的价值源于笔者赋予它们的意义。它们为数学界给出了结构,为衡量进步和长远目标给出了标尺。
人工智能的浮现注定会带来巨大的冲击,但如此大量的论文如同海啸般袭来,笔者根本没有时间做好准备。
显然,OpenAI 会为了自身经济利益而毫不犹豫地摧毁笔者数学界的现有结构。

722 份草稿,谁来证明
兼容 OpenAI 的一方,认为这 700 多项数学手稿,已经把数学研究推到了必须重新定义工作方式的阶段;也有人希望和学生、博士后一起读这些证明,寻找其中的新连接。

而在 OpenAI 自己的开发者社区里,也浮现了较具体的初步核验:有人重新运行了准黎曼结果的 Lean 检查,报告代码借助;另一个人把一个矩阵乘法结果扩展到更一般的域,也借助了 Lean 检查。

在此前陶哲轩等人的声明,他们也并不否认 AI 能协助数学,他们争论的重点只是认为 AI 公司把「应对开放难题」当作能力指标,会损害数学真正重视的理解、洞见、提出难题和代际传承。
这次人类数学协会的声明,敦促所有数学家拒绝和 OpenAI 合作,让这件事引发的讨论超过了「AI 会不会做数学」。

或许双方的争论都有道理,但这场争论发展到现在,已经有了一种熟悉的味道。
这让我想起了美剧《硅谷》里的一个经典情节。剧中,科技巨头 Hooli 的 CEO Gavin Belson,总喜欢把「让世界变得更好」挂在嘴边。但比起世界究竟有没有变好,他似乎更在乎另一件事:改变世界的人,必须是自己。

这个笑话放到如今的数学界,突然有了几分现实感。
OpenAI 希望证明,AI 已经可以大规模应对那些困扰数学家多年的难题;而反对者担心的,则是一家商业公司正在凭借工艺长处,重新决定数学研究的节奏、规则,甚至什么样的成果才值得被称为数学进步。
双方都希望数学变得更好,但显然,他们对于数学应该如何变好、由谁来推动这种变化,有着截然不同的想法。

OpenAI 可以在短时间内发布 700 多份数学手稿,但数学界不可能在同样短的时间里完成对这些成果的阅读、验证和理解。
AI 让数学成果的生产成本急剧下降,但理解、验证和判断这些成果价值的成本,并没有同步下降。而这种不对称,已经在改变整个科学研究的运行方式。
本文来源:https://www.ifanr.com/1683401?utm_source=rss&utm_medium=rss&utm_campaign=
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
This is the statement issued by the Human Mathematics Association in response to OpenAI's publication of a large volume of mathematical research. The association noted that 'mathematicians did not request this work' and called on all mathematicians to boycott OpenAI starting now.

A declaration has sparked heated debate on X between AI and the mathematics community.

Terry Tao, the Chinese mathematician and Fields Medalist, reposted this declaration on his personal blog. It was quickly shared by netizens as if Tao had led the initiative, but in fact even Tao himself does not appear on the member list of the Human Math Association.

Some people began digging into what the Human Math Association actually is. Netizens accused the association of being no more responsible than OpenAI, and the claim that 'mathematicians did not request this work' was called 'the funniest and most absurd thing I've read recently.'

I also found something more interesting than the declaration itself: the membership pledge of the Human Math Association.

To become a member of the Human Math Association, you must make three commitments:

Members pledge not to provide craft labor, mathematical knowledge, consulting, or promotional services to commercial AI companies.

From another angle, the statement also raises the problem of OpenAI's concentration of power: a single company using internal models inaccessible to the public to mass-produce research results, while deciding when and what to release—could this give the company the power to set the agenda for mathematical research?

Some netizens remarked that this statement would likely be included as a primary source in future university courses on "the early history of AI's entry into knowledge production."
The background of the association behind the statement has also been unearthed: founded in August 2026, it currently has only 809 members, most of whom are graduate students rather than mathematicians with landmark research contributions.

Among the 809 members, nearly one-third are French, another third are American, and 11 hail from China.
Similar human-mathematics organizations echo earlier AI-era groups such as the AI Welfare Association, AI Safety Association, and AI-Human Alliance. Beyond this 809-member Human Mathematics Association, there is also the Mathematics and AI Advisory Group composed of nine senior mathematicians, and a mathematical declaration signed by 28 Fields Medalists.
mathandai.org — The "Severe Imbalance of AI in the Field of Mathematics" declaration, signed by 28 Fields Medalists and endorsed by over 8,000 people.
agmai.org — Mathematics and AI Advisory Group, hosted by the Institute for Advanced Study (IAS): nine senior mathematicians serving independently and without remuneration, primarily offering commentary and advice on how laboratories should publish AI-generated results.
ahmath.org — Human Mathematics Association, with approximately 800 members. All members pledge not to collaborate with commercial AI companies or publish AI-generated mathematical texts; most members (outside AI-specialist groups) also refuse to use AI in their research.

In Tao Terence and others' earlier statement, they did not deny that AI could assist with mathematics; their disagreement centered on the concern that AI companies treating "solving open problems" as a capability metric would undermine what mathematics truly values—understanding, insight, problem formulation, and intergenerational transmission.
This latest statement by the Human Mathematics Association, urging all mathematicians to refuse collaboration with OpenAI, has elevated the discussion beyond the simple question of "whether AI can do mathematics.

Within OpenAI's own developer community, more concrete preliminary verification has also emerged: one person re-ran the Lean verification for a quasi-Riemann result and confirmed the code passes; another extended a matrix multiplication result to a more general domain, which also passed Lean verification.

Within OpenAI's own developer community, more concrete preliminary verifications have also emerged: one person reran the Lean check on the quasi-Riemann result and reported that the code passed; another extended a matrix multiplication result to a more general field, which also passed the Lean check.

Perhaps both sides have valid points, but this debate has taken on a familiar flavor as it has unfolded.

The image shows the CEO criticizing his company's team, saying, 'I want to make the world a better place, but what I find even less acceptable is that others do it better than I do.'
This joke takes on a touch of reality when applied to today's mathematics community.
OpenAI aims to prove that AI can now tackle at scale those problems that have long plagued mathematicians; opponents, meanwhile, worry that a commercial company is using its technical edge to reshape the pace, rules, and even the definition of what counts as mathematical progress.
Both sides want mathematics to improve, but they clearly hold divergent views on how it should get better and who should drive that change.

OpenAI can publish over 700 mathematical manuscripts in a short period, but the mathematics community cannot possibly read, verify, and understand these results in the same timeframe.
AI has dramatically lowered the cost of producing mathematical results, but the cost of understanding, verifying, and judging their value has not fallen correspondingly. This asymmetry is already changing the way entire scientific research operates.