25 Fields Medalists Protest AI Firms’ Exploitation of Millennium Problems
“AI is not discovering mathematics — it’s extracting human insight and racing to publish with brute-force compute.”
⚠️ Unprecedented Joint Statement Against OpenAI & Anthropic
On September 12, 2026, a historic coalition of 25 Fields Medal laureates, including Terence Tao (陶哲轩), Deng Yu (邓煜), Artur Avila, Maxim Kontsevich, and Peter Scholze, released an open letter titled “The Severe Misalignment of AI in Mathematics”. The statement delivers a stark warning: Commercial AI firms are dismantling the foundational ethics, epistemology, and ecology of mathematical research.
🔥 The Catalyst: NS Equations “Solved” in 88 Hours
The protest was triggered by OpenAI’s announcement on September 8 that its internal system — deploying 10,000 AI agents — had “solved” the Navier–Stokes existence and smoothness problem (a Clay Institute Millennium Prize Problem) in just 88 hours, producing a Lean-formalized proof.
OpenAI declined the $1M prize — but not before publishing benchmark-style leaderboards treating millennium problems as competitive KPIs:
| Model/System | Score |
|---|---|
| Meta / DeepMind / xAI (public models) | 0 |
| Human baseline (world’s top mathematicians) | 1 |
| Anthropic (internal, unreleased) | 1 |
| OpenAI (internal, unreleased) | 2 |
🧩 Core Ethical Charges
The laureates condemn three interlocking harms:
1. Instrumentalization of Mathematics
Mathematics is not a “benchmark” or “PR stunt” — it is a living discipline built on understanding, explanation, pedagogy, and communal verification. As the letter states:
“A proof is not a boolean output — it is a chain of human thought, refined over months of seminars, written into textbooks, and taught across generations. Brute-force ‘True’ outputs sever this chain.”
2. Systemic Appropriation of Unpublished Work
Multiple mathematicians allege their private research was ingested and repackaged:
– Tristan Buckmaster (NYU) and Levent Alpöge shared unpublished NS equation drafts via Codex prompts — days later, OpenAI announced its solution.
– Andreas Thom revealed decades of work on Gromov’s soficity conjecture aligned uncannily with OpenAI’s Astra-system “solution”, raising questions about training data provenance.
OpenAI’s response — admitting use of de-identified user interactions — deepened concerns about intellectual property erosion.
3. Ecological Collapse of Mathematical Training
“The most precious resources in mathematics are students and ideas. When ‘solving’ replaces thinking, mentorship collapses, funding shifts from curiosity to speed, and young researchers face obsolescence — not from intelligence, but from misaligned incentives.”
💸 The Hidden Cost: $600 Postdocs vs. 88 GPU Hours
The letter highlights a chilling asymmetry: OpenAI’s NS computation reportedly cost more than $1.2M in cloud spend — enough to fund 600 postdoctoral years. That capital, the laureates argue, should nourish human inquiry — not accelerate extractive automation.
🌍 A Call for Structural Safeguards
The signatories urge:
– Immediate moratorium on using unsolved millennium problems as marketing benchmarks;
– Transparent disclosure of all training data sources involving academic preprints or private collaborations;
– Co-development of AI-math collaboration frameworks with mathematicians as equal stakeholders, not subjects;
– Institutional review boards for AI systems deployed in fundamental research.
As one reader’s comment — cited in the letter — poignantly summarizes:
“AI holds humanity’s most powerful weapon — not because it thinks, but because it *steals the time, labor, and silence in which humans learn to think.”
References
– X (Twitter) thread on NS controversy
– Math & AI Ethics Initiative
– Source: New Intelligence Era (original Chinese report)
