OpenAI posts 722 AI-written math manuscripts; mathematicians push back

The central change is not that mathematics has been automated, but that producing plausible research at scale is becoming separable from validating it.

What this means. On October 6, OpenAI published 722 manuscripts grouped into 372 result families, generated predominantly by an unreleased internal model after approximately 4,000 problems were attempted. OpenAI estimates that the average result consumed compute equivalent to three hours of ChatGPT Pro thinking. The repository includes Lean formalizations for some results and abbreviated reasoning summaries: openai.com and github.com.

The initial catalogue was already down to 719 manuscripts after three withdrawals on October 7. A sign error invalidated an argument in one manuscript and a construction used by two dependent manuscripts. OpenAI also reported repairing 14 other manuscripts: github.com. This is a demonstration of both AI's research potential and the danger of treating publication volume as verified discovery.

Verification must become part of the result. Lean can mechanically check formalized proofs, but a manuscript without a formalization still requires mathematical scrutiny; formal verification also depends on correctly specifying the theorem and relevant assumptions. The repository's own documentation acknowledges that some unformalized results may contain issues: github.com and lean-lang.org.

The independent Advisory Group on Mathematics and Artificial Intelligence explicitly says that publication begins, rather than completes, the process of incorporating AI-generated work into mathematical knowledge. It also argues that mathematicians must retain the freedom to formulate their own questions and that access to research tools and computational resources matters: agmai.org.

Credit and incentives. The immediate risk is that researchers spend months independently pursuing problems already addressed in inaccessible AI experiments, or that their contributions become difficult to distinguish from machine-generated work. Reporting on the release describes concerns about disrupted research programmes and early-career mathematicians: businessinsider.com. Mathematical credit should distinguish who formulated a problem, who developed the argument, who verified it, and who wrote or revised the exposition. AI systems cannot take scholarly responsibility on behalf of human authors.

What should happen now?

• Mathematicians: Prioritize independent replication of consequential results, examine dependencies between manuscripts, publish corrections openly, and preserve dated records of contributions. Do not assume a result is either true or false simply because AI produced it.
• Students: Learn proof-writing and formal verification alongside AI-assisted exploration. Treat model output as a conjecture or candidate proof until checked.
• Journals: Require explicit AI-use disclosures, accountable human authorship, reproducible supporting materials, and clear distinctions between formally verified and independently reviewed claims. Establish rapid correction procedures for dependent results.
• AI labs: Release accessible models or equivalent research access, document prompts and methods sufficiently for replication, label verification status paper by paper, publish dependency maps, fund independent checking, and provide durable version histories.

The need for these safeguards is consistent with AGMAI's September recommendations, which explicitly ask AI labs to stop testing advanced mathematical problems on proprietary models unavailable to the broader scientific community: agmai.org.

What happens next? The most likely near-term outcome is a mixed research workflow: AI generates candidate results, humans determine significance and correctness, and formal proof assistants increasingly handle checkable portions. The decisive measure will not be how many manuscripts a model writes, but how many survive independent scrutiny and become useful mathematical knowledge.

1. openai.com
2. github.com
3. github.com
4. github.com
5. agmai.org
6. agmai.org
7. businessinsider.com
8. unite.ai
9. lean-lang.org
10. openai.com
11. developers.openai.com

What happened. On October 6, 2026, OpenAI posted 722 manuscripts in 372 families to github.com/openai/math, under the Apache-2.0 licence (unite.ai). They came from an unnamed internal model that was given about 4,000 problems. The average result used about three hours of ChatGPT Pro-level compute (github.com). About 42% of the top-line results are formalized in Lean (github.com), and about half were released unconfirmed (retractionwatch.com). On October 7, OpenAI withdrew three papers because a +1 should have been a -1. The error broke 'Algebraicity of Weil classes on split abelian eightfolds' and two papers that depended on it. OpenAI also revised 14 other manuscripts (retractionwatch.com, implicator.ai). The catalogue now lists 719 manuscripts (github.com).

What it means. 1) The hard part of mathematics is now checking results, not producing them. A batch this size outruns human refereeing. The withdrawals show that results not checked by computer can fail, and that versioned errata on GitHub become part of the record (github.com). Lean proves only the statement as it was typed in. A person still has to confirm that the formal statement matches the real conjecture (letsdatascience.com). 2) Outsiders can't reproduce the work. The prompts, the model's name and the compute per problem stayed private (letsdatascience.com). MIT's Andrew Sutherland said claims about one-shotting problems should be treated as unverified until the model is released (letsdatascience.com). 3) Credit is contested. The README gives only BibTeX citations and has no authorship policy (github.com). A credit dispute already followed OpenAI's Navier–Stokes claim in September (letsdatascience.com). 4) The community is split. The IAS-linked advisory group (AGMAI) asked labs to deposit results in archives no lab controls, to disclose prompts, time and cost, and to report failed attempts. It also said 'we do not endorse' testing on inaccessible models (unite.ai). The Association for Human Mathematics called the release 'a demonstration of power' and urged mathematicians to stop working with OpenAI (retractionwatch.com).

What comes next. Expect more errata, more Lean proofs and probably more withdrawals. OpenAI has promised all three (retractionwatch.com). It also says it is working to 'responsibly release' the model, but it has given no date (unite.ai). Reports say the model is still in training (aiweekly.co). Other labs will probably make similar mass claims, and the fights over release rules will get louder.

What to do now.

• Mathematicians: pick results in your own field and audit them, or formalize them in Lean. Each confirmed or broken result is real data. Credit humans who fix or verify a proof. Many of these problems are now 'claimed', not closed.
• Students: use the repository to learn. Check whether each Lean statement really matches the theorem, read the reasoning summaries, and practise finding errors. Being able to verify proofs has become a valuable skill.
• Journals and arXiv: require the model name, prompts, compute and a formalization status for any AI-produced result. Fund verification as a refereeing track, and keep withdrawn versions visible.
• AI labs: follow AGMAI in full. That means a neutral archive with DOIs, prompts and compute for each result, failed attempts, and selection criteria. Give independent researchers access so they can replicate the work, and fund the human verification your releases create (unite.ai).

1. unite.ai
2. github.com
3. retractionwatch.com
4. implicator.ai
5. letsdatascience.com
6. aiweekly.co

On 6 October 2026 OpenAI published a public GitHub repository (openai/math) containing 722 mathematical manuscripts (later reduced to 719) organised into 372 result families, produced by one unreleased internal frontier model. The model had been posed roughly 4,000 problems; each accepted result used, on average, the compute equivalent of about three hours of ChatGPT Pro thinking. Only a minority of the main results were accompanied by Lean formalisations (initially ~162 manuscripts; later reported as ~42 % of top-line results). OpenAI warned that some unformalised results could contain issues and committed to fixing them quickly. A day later it withdrew three manuscripts after a sign error invalidated a core argument and two dependent papers, while revising fourteen others. The release was framed as model-evaluation work, informed by consultations with the independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI) hosted at the Institute for Advanced Study, though AGMAI later stated that its role was not an endorsement of the process or a judgement of the results’ impact.

The episode demonstrates that frontier models can generate large volumes of candidate results on open problems in mathematics, theoretical computer science and related fields in a short time. It also exposes the gap between generation and the traditional processes of verification, exposition, citation, contextualisation and credit. Lean provides machine-checkable guarantees for the formalised subset, yet even formalised proofs require human scrutiny to confirm that the Lean statement matches the claimed mathematical result and that the surrounding exposition is intelligible. The rapid appearance of an error, the incomplete formalisation rate, and the sheer scale produced widespread concern that the community was being asked to sort “slop” or “flooding the zone.” Critics, including the Association for Human Mathematics (whose statement was amplified by Terence Tao), described the simultaneous release of more than 700 files as a demonstration of power rather than scholarship and urged mathematicians to stop collaborating with OpenAI. AGMAI and others stressed that public release is only a first step toward human understanding.

In the near term the community will continue auditing the repository: more Lean formalisations will appear, further errors or revisions will surface, and selected results will be absorbed or discarded. OpenAI has said it will fund workshops and improve future exposition, citations and disclosure. Journals, preprint servers and hiring committees will have to decide how to treat AI-generated manuscripts that lack clear human authorship, full formalisation or conventional literature context. Students and researchers will face both new tools and new risks of investing effort in results that later prove flawed or uninterpretable.

Mathematicians should prioritise independent verification of high-value claims, demand transparent problem-selection and failure data, and insist that formalisation status and natural-language/Lean correspondence be clearly labelled. Students should treat the manuscripts as unverified drafts until community consensus forms. Journals should develop explicit policies for AI-generated work (authorship, reproducibility, formalisation requirements). AI labs should release models or enough tooling for independent re-evaluation, publish negative results, and align release protocols more closely with community guidelines before generating further large batches.

1. github.com
2. openai.com
3. unite.ai
4. github.com
5. retractionwatch.com
6. nature.com
7. theverge.com
8. scientificamerican.com
9. agmai.org
10. ahmath.org
11. openai.com

On October 6 2026 OpenAI published 722 manuscripts organized into 372 families to public GitHub repository openai/math under Apache-2.0, produced by an unreleased internal frontier model. The model was posed roughly 4,000 problems and each result used on average about three hours of ChatGPT Pro thinking compute. Many manuscripts have Lean formalizations but only 162 had main results formalized at release, and the README warns some unformalized results could have issues and that verification is staged.

On October 7 the repository history recorded three withdrawals after a sign error in 'Algebraicity of Weil classes on split abelian eightfolds' invalidated a stabilization-trace cancellation argument and the construction used by two dependent papers. Those three are now withdrawn with gap notices. Fourteen other manuscripts were revised with proof repairs and 13 more were updated to cite revised editions, and the catalogue now lists 719 manuscripts in 372 families with about 42% of top-line results formalized (300/719).

Reactions were sharp. Coverage notes mathematicians balked at the scale, a group letter called releasing over 700 files at once 'not a demonstration of scholarship, but a demonstration of power.' The Association for Human Mathematics (AHM), chaired by Fields Medalist Terence Tao, accused OpenAI of violating core norms and urged mathematicians to stop working with it, arguing mathematicians did not ask for this work and that mass release burdens verification.

For mathematics this means theorem production is decoupled from human understanding for the first time at scale. The scarce resource shifts from solving to checking, explaining and integrating. Lean moves from optional supplement to necessary epistemic infrastructure, and results now exist in three states: machine-checked, human-readable but unformalized, and research lead.

How results are produced, checked and credited must change. Production is now automated curation: pose thousands of problems, filter by significance into families. Checking needs versioned public history, persistent identifiers and tiered status badges, not binary publish. Credit is unclear: BibTeX blocks list manuscript but not prompt, cost or human editors. The Advisory Group on Mathematics and AI at IAS, consulted by OpenAI, recommended labs deposit in repositories no lab controls, publish model name, prompts, summarized chain of thought, time and cost, formalize as far as possible, explain selection and failure rates, not use releases as marketing, and fund workshops, summer schools and expository books through nonprofits with equitable model access.

Next is likely: more withdrawals and revisions, community sprint to raise Lean coverage, explanatory rewrites in human style like the Anthropic model where researchers were paid to translate an AI key idea, and funded workshops OpenAI says it will run.

What to do now: mathematicians — triage by family and Lean status, avoid racing to claim, organize reading groups to turn alien proofs into teachable ideas; students — learn Lean 4 and how to interrogate reasoning summaries, but keep classical training for intuition; journals and arXiv — require disclosure of AI provenance, compute and verification level, keep version history accessible, protect against mass scraping for training; AI labs — follow AGMAI, publish prompts and abridged reasoning, share equitable access, fund understanding work, avoid 700-file surprise drops and coordinate responsible disclosure.

1. unite.ai
2. github.com
3. github.com
4. aiweekly.co
5. glonce.com
6. davidborish.com

OpenAI's release of 722 AI-generated math manuscripts (organized into 372 families) on GitHub shifts mathematical research from a human-paced endeavor of creation to a massive bottleneck of verification. Because an unreleased internal model produced the work and only a portion is checked in Lean, mathematicians face an unprecedented volume of unverified preprints. This challenges traditional notions of credit, peer review, and human intellectual ownership.

For mathematics, this means a flood of potential discoveries that risk overwhelming human vetting capacity unless paired with formal proof assistants. For production and credit, authorship attribution becomes murky, raising questions about how to credit an unreleased proprietary system.

What happens next will likely involve a scramble by independent mathematicians to formally verify or refute the remaining preprints, alongside growing institutional pushback against proprietary gatekeeping of advanced mathematical reasoning.

Mathematicians and students should leverage formal proof languages like Lean to independently verify machine outputs. Journals must tighten standards for AI-assisted submissions, requiring formal verification where feasible. AI labs should adopt open science practices by sharing underlying models or partnering transparently before flooding public repositories with unchecked technical claims.

1. unite.ai
2. github.com
3. retractionwatch.com