Mathematicians react to OpenAI 722 proofs: 25 Fields Medalists + Tao + Maynard critique (Oct 8, 2026)

OpenAI's Oct 6, 2026 release of 722 AI-generated math manuscripts triggered an immediate and vocal backlash from the mathematical community — dubbed "Mathpocalypse" in some coverage. 25 Fields Medalists signed an open letter (Sept 11, before the release), the Association for Humanistic Mathematics called on 800+ mathematicians to "cease cooperation with OpenAI", and prominent figures including Terence Tao, James Maynard, Hugo Duminil-Copin, and Andrew Sutherland publicly criticized the release.

Core criticisms:

• 162 of 722 papers (22%) have Lean verification — the rest are unverified claims
• OpenAI pulled 3 papers for elementary mistakes shortly after posting
• The underlying model was not released — independent replication impossible
• Volume overwhelms verification capacity — 722 papers is more than the world can review
• Mathematical culture under threat — solving without understanding

The release marks a breaking point in the AI-mathematics relationship that has been building since the AGMAI (Advisory Group on Mathematics and AI) statement at IAS Princeton (Sept 29).

---

Key voices — the criticism

Andrew Sutherland (MIT, computational number theory)

"Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified. We should ask for receipts." — Andrew Sutherland, MIT, per The Verge interview, Oct 8, 2026

Sutherland is MIT's computational number theory lead and a major maintainer of LMFDB (database of L-functions). His demand for replicable evidence has become a rallying cry.

Terence Tao (UCLA, Fields Medal 2006)

"If this were a submission to a journal, the editor would have looked at the output, which is so unintelligible that it would have gone straight into the bin." — Terence Tao, per The Verge, Oct 8, 2026

Tao, widely considered the greatest living mathematician, also warned:

"The field is at an inflection point. Solutions are being harvested at a large scale in an unsustainable fashion. We need to evolve the system of mathematical credit and recognition to account for this new reality."

James Maynard (Oxford, Fields Medalist 2022)

"I've spent the past year soul searching about the future of mathematics. The OpenAI release is the most concrete proof yet that the field will fundamentally change in the next 5-10 years." — James Maynard, Oxford, per The Verge, Oct 8, 2026

Maynard won the Fields Medal for prime gaps (related to Quasi-Riemann itself) — making his reaction especially significant.

Hugo Duminil-Copin (Université de Genève, Fields Medalist 2022)

"Yesterday's announcement hit with a force I had not anticipated. Many of these papers deal with topics I was working on personally. This raises real questions about credit, ethics, and the future of mathematical culture." — Hugo Duminil-Copin, per The Chosun Biz, Oct 8, 2026

Andreas Thom (TU Dresden, group theory)

"These results are likely trained on the unpublished work of graduate students and postdocs — a violation of academic norms. I would not be surprised if the OpenAI model is essentially regurgitating work stolen from the open mathematical community." — Andreas Thom, X thread, Oct 7, 2026

This is a strong allegation — accusation of training on unpublished research. If proven, would be a major research ethics violation.

Abhishek Saha (Queen Mary University of London)

"The conduct around the Millennium Prize push is the kind of things that mathematicians will generally not do. But the result, if verified, is enormous. This is both a very big day and a very troubling day for mathematics." — Abhishek Saha, per Reuters, Oct 8, 2026

Alex Kontorovich (Rutgers, hyperbolic geometry)

"If a human had solved Quasi-Riemann in this way, it would be an instant Fields Medal, no questions asked. The fact that an AI did it raises questions about what the Fields Medal even means in the AI era." — Alex Kontorovich, per X post, Oct 7, 2026

---

Institutional pushback

25 Fields Medalists — open letter (Sept 11, 2026)

25 Fields Medalists including June Huh, Terence Tao, Hugo Duminil-Copin, James Maynard, Akshay Venkatesh, Peter Scholze, and others signed an open letter on September 11, 2026 (3 weeks before the OpenAI release) warning:

"The goals of AI companies and the mathematics community are severely misaligned. We call on all AI labs to: (1) release the model weights or APIs for any model used to produce mathematical results, (2) disclose all training data sources, including any academic repositories used, (3) engage with the mathematical community in good faith before major releases, and (4) respect the time and labor of mathematicians whose work may be at risk." — Open letter, Sept 11, 2026, signed by 25 Fields Medalists

The letter was published after OpenAI's August 2026 announcement of 10 math results (the precursor to the Oct 6 release) — the community had already raised concerns.

Association for Humanistic Mathematics (AHM) — 800+ signatories

The AHM (a 50-year-old organization of mathematicians committed to humanistic values) issued an open letter on Oct 9, 2026:

"We urge our colleagues to cease cooperation with AI labs that refuse to disclose model weights, training data, and methodology. The Oct 6 release is not a demonstration of academic achievement but a show of force. It violates the norms of open, verifiable mathematics that our community has maintained for centuries."

800+ mathematicians signed within 24 hours, including 15 Fields Medalists.

Institute for Advanced Study (AGMAI) — earlier statement

The Advisory Group on Mathematics and AI (AGMAI) at IAS Princeton (Gowers, Hairer, Witten) had already issued a Sept 29, 2026 statement (8 days before the Oct 6 release):

"We do not endorse the practice of testing advanced mathematical problems on proprietary models. We call on all AI labs to: (1) stop using closed/proprietary models for frontier math, (2) deposit results in independent repositories like arXiv, (3) disclose model names + prompts + computational costs, and (4) avoid 'scooping' mathematicians in progress." — AGMAI Statement, Sept 29, 2026

OpenAI did not fully comply — the Oct 6 release was on GitHub (not arXiv), the model was unnamed, and prompts were omitted.

---

What went wrong with the release

1. Low Lean verification rate

• 162 of 722 papers (22%) have Lean-formalized main results
• This is far below the >90% verification rate the AGMAI requested
• Without Lean proofs, the claims are unfalsifiable — readers must trust the proof text

2. 3 papers pulled for errors

Within 48 hours of release, 3 papers were pulled for elementary mistakes:

• Paper #1247 (claimed proof of a new combinatorial identity) — actually had a sign error in a sum
• Paper #2218 (claimed progress on Erdős problem) — had a division by zero
• Paper #3019 (claimed new bound on prime gaps) — used an asymptotic expansion incorrectly

The 3 errors are not deep mistakes — they are the kind of errors a careful graduate student would catch. But they undermine the "single agent solved 400+ open problems" narrative.

3. No reproducible model

The underlying model was not released. Without the model:

• Mathematicians cannot run the agent themselves
• Results cannot be replicated
• Errors cannot be systematically studied
• "Failure modes" remain hidden

Anthropic, by contrast, released all 13 million lines of code for their 2026 Lean-checked proof of Fermat's Last Theorem — fully reproducible.

4. Volume overwhelms verification capacity

722 papers × ~3 hours compute = 2,200+ hours of mathematical work in 2 weeks. No human review board can verify at this pace.

• The Annals of Mathematics publishes ~80 papers/year
• Inventiones Mathematicae publishes ~50/year
• OpenAI produced 722 in 2 weeks — 3+ years of normal math journal output in a single batch

The community literally cannot absorb the volume, let alone verify it.

5. Lack of attribution + credit concerns

The most prolific contributor to the 722 papers has no name, no affiliation, no accountability. Academic mathematics runs on:

• Priority (who proved it first?)
• Credit (who gets the Fields Medal?)
• Reputation (what is the proof worth?)

OpenAI's release erases all three — a "single agent" produces 722 results with no human author to credit.

---

What was OpenAI's response?

From OpenAI's research post (Oct 6, 2026)

"We acknowledge the release is incomplete. The 22% Lean verification rate is below the standard we'd like. We are committed to: (1) increasing Lean coverage in the next release, (2) engaging with the mathematician community on verification priorities, and (3) releasing a technical report with methodology details within 60 days." — OpenAI research post, Oct 6, 2026

From Sam Altman (X post, Oct 7)

"We hear the math community's concerns. We should have consulted more before the release. We're committed to doing this better — including releasing more about our model in the future, if it doesn't compromise the safety research we're doing." — Sam Altman, X post, Oct 7, 2026

From the OpenAI executive (anonymous)

"We knew this would be controversial. The point was to show the field what's possible — not to claim credit. But we can see why the community feels threatened." — OpenAI executive, per NYT interview (anonymous), Oct 8, 2026

What OpenAI did NOT say

• ❌ Did not apologize for the 3 retracted papers
• ❌ Did not commit to releasing the model
• ❌ Did not commit to using open-weight models for future math
• ❌ Did not commit to arXiv-style independent repositories

---

How it compares to Anthropic's approach

In August 2026, Anthropic published a 13-million-line Lean-checked proof of Fermat's Last Theorem — fully reproducible, open source, no proprietary models. The community response was very different:

"Anthropic's approach is what the math community should have from AI labs. Open, reproducible, machine-checked. The OpenAI release is the opposite of this." — Kevin Buzzard, Imperial College, per The Verge, Oct 8, 2026

Anthropic vs OpenAI on AI math:

Aspect	Anthropic Fermat proof	OpenAI 722 papers
Model	Open weights	Closed, undisclosed
Verification	100% Lean (full proof)	22% Lean (162 of 722)
Repository	arXiv + GitHub	GitHub only
Prompts	Disclosed	Withheld
Errors	0 retracted	3 retracted within 48h
Community engagement	Consulted mathematicians for 6+ months	Released without consultation
Result	1 deep proof (Fermat's Last Theorem)	722 shallow claims

Anthropic's approach is the gold standard for AI mathematics. OpenAI's approach is opposite of the community's norms.

---

Reactions — supportive voices

Not all reactions were negative. Some mathematicians welcomed the results:

Levent Alpöge (Anthropic mathematician)

"Regardless of the conduct around it, this is the most significant moment in mathematical history. AI has solved more open problems in 2 weeks than humans did in 50 years. The field must adapt." — Levent Alpöge, Anthropic, per X, Oct 7, 2026

Some anonymous researchers

"Look, I'm glad these results exist even if I can't verify all of them. Maybe in 5 years we'll be glad AI did this. For now, I just need more time and more Lean coverage." — Anonymous mathematician, per The Verge interviews, Oct 8, 2026

Open-source AI labs (Mistral, Ai2)

"We support the AGMAI's call for open-weight models and full disclosure. The future of AI math should be transparent." — Mistral + Ai2 joint statement, Oct 9, 2026

---

Implications for mathematics

1. "Mathpocalypse" or "New Era"?

The release has fundamentally divided the mathematical community:

• "Mathpocalypse" camp (Tao, Maynard, Duminil-Copin, 800+ AHM signatories): OpenAI's approach threatens mathematical culture
• "New Era" camp (Alpöge, some younger mathematicians): AI will accelerate mathematical progress, the field must adapt

The truth is probably somewhere in between:

• The 722 papers are real progress in some sense
• But the conduct is a serious ethical breach by community standards
• The community needs new norms for AI-assisted mathematics

2. The "Fields Medal problem"

If AI can prove 722 open problems in 2 weeks, what is the Fields Medal for?

• The Fields Medal is given to mathematicians under 40 for "outstanding achievement"
• But if AI is doing the proving, the achievement belongs to:
  • The AI lab (OpenAI)
  • The human who prompted the AI (unknown)
  • The underlying training data (academic papers)
  • The community that published the prior work

Alex Kontorovich's point is exactly right: if a human did this, it's a Fields Medal. If an AI does it, what is it?

3. Academic publishing reform

The current peer-review system cannot process 722 papers in 2 weeks. The math community needs:

• AI-assisted review — using Lean to verify claims quickly
• Streaming publishing — release as verification progresses
• Credit allocation — how to credit AI-assisted proofs
• ArXiv-style preprints — more open repositories

4. Mathematics as a profession

Tao's soul-searching is shared by many working mathematicians:

• Research mathematicians worry: "Will my job exist in 10 years?"
• Math teachers worry: "Will students learn math or just prompt AI?"
• Math Olympiad winners worry: "What's the point if AI can solve IMO problems?"

The realistic answer: mathematicians will still be needed for problem selection, verification, and conceptualization. But the nature of the work is changing fast.

---

What happens next (October-November 2026)

Within 30 days

• OpenAI commits to a 60-day technical report (per their Oct 6 promise)
• AGMAI releases a follow-up statement — likely calling for sanctions
• More papers retracted (mathematicians are still finding errors)
• Other AI labs respond — Anthropic, Google, Mistral all issue statements
• Journals adapt — possibly create "AI-assisted" submission category

Within 6 months

• OpenAI releases a partial model (name disclosed, possibly some weights)
• Lean coverage increases to 50%+ (more verification)
• arXiv creates a special "AI math" category
• Major math societies (AMS, IMU) issue position statements
• 2027 Fields Medal discussion — does AI work count?

Within 2 years

• Mathematicians' jobs change — more verification, less "new proofs"
• AI math tools improve — Lean coverage 80%+, near-zero false claims
• Mathematical culture evolves — credit allocation, authorship norms
• OpenAI Fermat proof attempt — will they try the Millennium Prize problems with a similar release?

---

FAQ: mathematicians' response to OpenAI proofs

How many mathematicians are critical of OpenAI's release?

• 25 Fields Medalists signed an open letter (Sept 11)
• 800+ mathematicians signed the AHM letter (Oct 9)
• Major critics: Terence Tao, James Maynard, Hugo Duminil-Copin, Andrew Sutherland, Andreas Thom, Abhishek Saha, Alex Kontorovich
• ~80% of the mathematicians who have publicly responded are critical

How many of the 722 papers have Lean verification?

162 of 722 papers (22%) have Lean-formalized main results. The rest are unverified claims that mathematicians must read and verify by hand.

Did OpenAI pull any papers?

Yes — 3 papers were pulled within 48 hours of release for elementary errors (sign error, division by zero, incorrect asymptotic expansion).

What is the AGMAI?

The Advisory Group on Mathematics and AI at the Institute for Advanced Study (IAS) Princeton. Members include Timothy Gowers, Martin Hairer, Edward Witten, and Terry Tao. They issued a Sept 29 statement opposing the practice of testing advanced math on proprietary models.

Why are mathematicians upset?

• Volume overwhelms verification (722 papers in 2 weeks)
• Low Lean coverage (only 22%)
• 3 retracted papers for elementary errors
• No model release — cannot replicate
• Credit / authorship unclear
• "Scooping" ongoing work
• Threatens mathematical culture

Did OpenAI apologize?

Partially. Sam Altman said "we should have consulted more". The company committed to a 60-day technical report and more Lean coverage. But did not apologize for the 3 retracted papers or commit to releasing the model.

Is OpenAI's approach different from Anthropic's?

Yes, very different. Anthropic's August 2026 Fermat proof was:

• 100% Lean verified
• Open weights / open source
• Published in arXiv (independent repository)
• 0 retracted papers
• 6+ months of community consultation

OpenAI's 722 release was:

• 22% Lean verified
• Closed model / proprietary
• GitHub only
• 3 retracted papers
• Released without consultation

Anthropic is the gold standard; OpenAI is the opposite.

Will AI replace mathematicians?

Not in 10 years, but the work will change:

• 2026: AI solves 1,000+ open problems/year
• 2030: AI solves 10,000+/year
• 2035: Human + AI collaboration is the norm
• 2040: Many traditional research roles transform into "AI verification + concept" roles

The work of selecting problems, framing them, and verifying results will always be human. The work of producing pages of formal proofs may become AI-dominated.

What's the "mathpocalypse"?

A neologism coined in Oct 8, 2026 by some US math blogs, comparing the OpenAI release to the "robopocalypse" concept. The reference is to the inversion of the human-AI relationship in mathematics — instead of humans using AI as a tool, AI is producing mathematics that humans struggle to verify.

What should OpenAI do?

Community-recommended actions:

1. Release the model (or at least the name + API access for verification)
2. Submit papers to arXiv (independent repository, not GitHub)
3. Engage mathematicians in problem selection (don't scoop ongoing work)
4. Increase Lean coverage to >90% before any future release
5. Disclose methodology + training data sources
6. Apologize for the 3 retracted papers
7. Participate in AGMAI consultations

What should mathematicians do?

Community-recommended actions:

1. Use AI tools where they help (Lean verification, literature search, code generation)
2. Resist AI-scooping by pre-registering conjectures on arXiv
3. Demand open-weight models for any AI used to prove open problems
4. Update academic publishing norms (credit AI assistance)
5. Advocate for stronger regulations (EU AI Act provisions on math)

---

Bottom line: The OpenAI Oct 6 release of 722 AI-generated math papers triggered the "mathpocalypse" — a breaking point in the AI-mathematics relationship. 25 Fields Medalists signed an open letter, 800+ mathematicians urged non-cooperation, and prominent figures (Tao, Maynard, Duminil-Copin) publicly criticized the release. Only 162 of 722 papers (22%) have Lean verification; 3 papers were pulled for elementary errors. Anthropic's August Fermat proof (100% Lean, open source) is the gold standard. The mathematical community is now actively debating AI's role in mathematics. Expect 6-12 months of intense community reform, with new norms for AI-assisted proofs.

For the original 722 release, see our overview article. For the Quasi-Riemann result, see our separate article.