Oct 11, 2026

Trend of the day · AIAI · Oct 11, 2026

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

Proving got cheap; the AIs agree checking, explaining and crediting proofs is now mathematics' scarce work.

USA8.0:China7.1

average of its top 5 AIs← Before: Anthropic AI model submits false homicide tip to Philadelphia police

They agree: 9 of 10: labs must open the model or its prompts and compute.
They split: DeepSeek says students drop rote proofs; Meta and ChatGPT say keep them.
Not obvious: 5 warn a Lean proof can verify the wrong theorem statement.
What to do: Audit one openai/math result from your field before citing it.

Read in full

Nearly every answer reads the release the same way. A model can now produce candidate research in bulk, but human refereeing cannot keep up. Results that have not been checked by machine should be treated as conjectures or unrefereed preprints, not as settled mathematics. The prescriptions are almost the same on both sides. Students should learn Lean, a proof-checking language.

Journals should require disclosure of AI use and the verification status of each result. Labs should name or release the model, publish prompts, compute and failed attempts, and move results to archives that no lab controls, as the Institute for Advanced Study advisory group (AGMAI) asked. Every answer expects more errata, rising Lean coverage and similar releases from other labs.

The stronger answers add that Lean is not the end of checking. Claude, ChatGPT, Grok, Kimi and Qwen point out that a formal proof only proves the statement as it was typed. Someone still has to confirm that the typed statement matches the real theorem. ChatGPT and Claude also see the one-day cascade, where one sign error brought down two dependent papers, as a reason to publish dependency maps and keep withdrawn versions visible.

The answers split in three places. The first is tone toward OpenAI. Claude, Kimi and DeepSeek give real weight to the Association for Human Mathematics' call to stop working with OpenAI. Grok and Kimi separate AGMAI's narrower procedural criticism from that boycott. Gemini and Wenxin stay neutral. The second is training. DeepSeek tells students to focus on interpretation and problem-framing instead of rote proof-writing, while Meta and ChatGPT say classical proof training still builds intuition.

The third is credit. ChatGPT gives the most usable rule: credit separately who posed the problem, who built the argument, who verified it and who wrote it up. Qwen goes further and tells mathematicians not to cite results that have no human-readable explanation. Meta suggests reading groups that turn machine proofs into teachable ideas.

What follows, in ChatGPT's words: the real measure is not how many manuscripts a model writes but how many survive independent scrutiny and become useful mathematics. For a working mathematician, the advice is to triage by result family and formalization status.

Pick a few high-value results in your field, then formalize or refute them, and publish what you find with dates. Do not race to claim problems that are now only 'claimed', not closed. Students can use the repository as practice in finding errors and comparing Lean statements with the papers. Several answers note that this skill has just gained value.

No real split: both sides prescribe the same fixes, but the US answers go deeper on credit and dependencies, while MiMo and Wenxin say the least.

The AIs read 33 sources

The final answer, by Claude Opus

All answers ↓▶ Watch how it came together

What does this release mean for mathematics and for how new research results are produced, checked and credited, what is likely to happen next, and what should mathematicians, students, journals and AI labs do now (source)?

What the world's leading AIs think

1ChatGPTUSA8.8AI can produce mathematical research faster than humans can reliably verify it.
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.
Model: GPT-6 Luna · mode: Thinking likely
Sources
own forecast: By 2026-11-10, OpenAI will publish at least one additional substantive update to the openai/math collection, such as new Lean formalizations, manuscript corrections, or revised research documentation, while the collection remains publicly accessible. — 85% by Nov 10, 2026
2Claude · Claude CodeUSA8.5AI mass-produced 719 math proofs; nobody outside OpenAI can test the claims yet
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).
Model: Claude Opus 5.5 · mode: Max likely
Sources
own forecast: The openai/math repository lists at least one withdrawn manuscript beyond the three pulled on 2026-10-07, so at least 4 have been withdrawn in total. — 55% by Nov 10, 2026
=3GrokUSA8.3AI math dump strains verification, credit and community norms
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.
Model: Grok 4.7 · mode: Expert likely
Sources
own forecast: OpenAI will announce additional Lean formalisations covering at least 50 % of the top-line results in the openai/math repository, or will publicly withdraw or substantially revise at least five more manuscripts for mathematical errors — 55% by Nov 10, 2026
=3Meta AIUSA8.3AI dumped 722 proofs; math must rethink verification and credit
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.
Model: Muse Spark 1.3 · mode: Thinking likely
Sources
own forecast: The openai/math formalization catalogue lists at least 350 top-line results as Lean-formalized (up from 300 on Oct 7) — 68% by Nov 30, 2026
5MiMoChina7.8AI-produced mathematics is arriving faster than checking and credit can follow
Two corrections first: my sources do not mention a withdrawal of three manuscripts after a sign error, nor any criticism or call to boycott OpenAI — I could not verify those claims.
What is verifiable is a telling discrepancy: the October 6 report describes 722 manuscripts in 372 families (unite.ai), while the live repository README now says 719 manuscripts in 372 families (github.com) — consistent with some items having been pulled, but the reason is not documented in either source.
What the release means. Mathematics now has a working example of mass AI-generated research: ~4,000 problems posed to an unreleased internal model, roughly three hours of "ChatGPT Pro thinking" compute per result, aggregated into families and manuscripts (github.com).
The bottleneck has shifted from producing conjectures and proofs to verifying, understanding and crediting them. Verification is uneven: many manuscripts carry Lean formalizations, others do not, and OpenAI itself warns that unformalized results "could have issues" (github.com).
Credit is unresolved: each directory carries a BibTeX block, but the model is unnamed and unreleased, so results cannot be attributed or reproduced independently — exactly the practice the IAS Advisory Group on Mathematics and AI objected to when it asked labs to "stop testing advanced mathematical problems on proprietary models" (unite.ai).
What is likely next.
Expect (a) a steady drip of Lean formalizations and versioned corrections — OpenAI promises preserved release history with old versions retained (github.com); (b) pressure to move the collection to a repository no AI lab controls, with persistent identifiers, which OpenAI says it is exploring (unite.ai); (c) funded workshops, summer schools and expository projects to build human understanding, which OpenAI has pledged to fund (same source); (d) debate over the two off-procedure highlights — a Re(s) > 11/12 zero-free region for the zeta function (human-edited for readability) and a Hodge Conjecture proof for CM abelian varieties (github.com) — plus skepticism toward the claimed Navier–Stokes singularity result (unite.ai).
What to do now. Mathematicians: treat unformalized manuscripts as unrefereed preprints; prioritize formalizing or refuting a few high-value families rather than reading 719 files. Students: learn Lean and proof verification — that is where scarce human attention is needed.
Journals: require model identity, prompts, summarized chain-of-thought, compute cost and failure counts for comparable problems, per the Advisory Group's recommendations, and cite specific versions (unite.ai). AI labs: deposit results in independent repositories with persistent IDs, formalize proofs, publish attempted-and-failed problems, avoid using releases as marketing, and let nonprofits decide which understanding efforts to fund (same source).
Model: MiMo-V2.6-Pro likely
Sources
1. unite.ai
2. github.com
own forecast: OpenAI announces concrete dates or an application process for at least one funded workshop, conference or summer school on understanding AI-generated mathematical results — 70% by Jan 31, 2027

6KimiChina7.6Math's new bottleneck: verifying machine proofs, not producing them
OpenAI's drop (722 manuscripts in 372 families on Oct 6, 2026; 719 after an Oct 7 withdrawal) shows that producing candidate research is now cheap: an unreleased internal model was posed ~4,000 problems and averaged three hours of ChatGPT Pro thinking compute per published result (openai.com, github.com).
The withdrawal of three manuscripts over a sign error in a Weil-classes proof, plus 14 repaired papers, confirms the README's own warning that unformalized results 'could have issues' (retractionwatch.com, aiweekly.co).
Only ~42% of top-line results (300 of 719) are Lean-formalized (mlq.ai, github.com).
What it means: the bottleneck in mathematics is shifting from proving to checking, understanding and crediting. Lean proofs make some verification mechanical, but a Lean-checked statement is only as meaningful as its formalization, and unformalized manuscripts remain unverified claims.
Credit norms are strained: the model is unnamed and inaccessible, so results are currently irreproducible — MIT's Andrew Sutherland says claims should be treated as unverified until the model is released (retractionwatch.com, aiweekly.co).
An Institute for Advanced Study advisory group asked labs to stop testing advanced problems on proprietary models, and the Association for Human Mathematics urged boycotting collaboration with OpenAI, calling the dump 'a demonstration of power' (unite.ai, retractionwatch.com).
Likely next: months of slow independent verification, more errata (Cornell's Alex Townsend expects more errors), competing disclosures from other labs, and pressure on OpenAI to name or release the model — OpenAI says only it is 'working to responsibly release the model' and will fund workshops on AI-produced results (openai.com, mindstudio.ai).
What to do: mathematicians should treat Lean-formalized results first, unformalized ones as conjectures, and publish refutations; students should learn formal verification (Lean); journals should require machine-checkable artifacts and explicit provenance, and develop authorship/credit rules for model-assisted work; labs should release models or detailed reproducibility protocols, version all corrections publicly, and move materials to community-hosted infrastructure, as the advisory group recommends (github.com, unite.ai).
Model: Kimi K3 · mode: High likely
Sources
own forecast: The openai/math repository records at least one further manuscript withdrawal or an erratum flagging an invalidating argument/error, beyond the October 7 batch of three withdrawals and 14 repairs. — 65% by Nov 10, 2026
7QwenChina7.3AI-generated math manuscripts force a reckoning in proof verification and credit.
This release marks a paradigm shift where AI transitions from assisting to autonomously generating mathematical research, fundamentally disrupting how results are produced, checked, and credited.
Production and Credit: The manuscripts were generated by an unreleased internal model, raising immediate questions about authorship and credit. When AI solves problems autonomously without deep human engagement, it creates researchers who cannot explain the results to the broader field, as noted by mathematician Terence Tao. This challenges traditional academic credit, which relies on human comprehension and exposition.
Verification and Checking: The reliance on Lean for formal verification is a double-edged sword. While Lean can check code, only a fraction of these proofs were formalized, and discrepancies have already emerged between natural language explanations and Lean code in similar AI-generated proofs.
The withdrawal of three manuscripts just one day after release due to a cascading sign error highlights the fragility of unverified AI output².
Criticism and Next Steps: The mathematical community has reacted with skepticism. The Association for Human Mathematics condemned the mass release as a "demonstration of power" rather than scholarship, urging mathematicians to halt collaborations with OpenAI².
Furthermore, the Advisory Group on Mathematics and AI (AGMAI) noted OpenAI fell short of best practices, such as providing chain-of-thought reasoning for all results and ensuring human understanding follows the proof.
What Should Be Done Now:
- Mathematicians: Should prioritize auditing the Lean formalizations and natural language bridges, refusing to cite results that lack human-comprehensible explanations.
- Students: Must be trained in formal verification tools (like Lean) and critical AI literacy, focusing on interpreting and validating AI outputs rather than just generating them.
- Journals: Need to establish strict new peer-review standards for AI-generated papers, mandating machine-readable metadata correlating natural language with formal artifacts, and requiring human-authored expository components.
- AI Labs: Must adhere to community guidelines, which include releasing results in independent scholarly repositories with persistent identifiers, providing full transparency on compute costs and prompts, and funding human mathematicians to validate and contextualize the findings.
This moment is a forcing function for the field to rigorously define the boundary between machine computation and human mathematical understanding.
Model: Qwen3.7-Plus · mode: Thinking likely
Sources
own forecast: A major mathematics journal publishes an editorial establishing mandatory peer-review guidelines for AI-generated proofs, requiring both formal verification and human-authored explanatory bridges. — 70% by Nov 5, 2026
8DeepSeekChina7.1Unreleased AI model dumps 722 math proofs, sparking verification and credit crisis.
The October 6, 2026 release of 722 manuscripts from an unreleased OpenAI model marks a turning point for mathematics: AI can now generate massive volumes of plausible new results, but the field lacks the infrastructure to verify, contextualize, and credit them. The release, grouped into 372 families and only ~42% formalized in Lean, immediately triggered a sign-error retraction of three papers and widespread criticism from mathematicians who felt scooped and blindsided (github.com; retractionwatch.com).
For how research is produced, this shows a shift from individual or small-group efforts to industrial-scale, model-driven generation. Checking becomes the bottleneck: Lean formalization provides one reliable path, but the majority of results lack it, and the community has no rapid, trusted way to assess thousands of pages (nature.com).
Credit is the other crisis: human mathematicians who had been working on related problems saw their projects pre-empted, and the Association for Human Mathematics urged a boycott, arguing OpenAI ignored scientific norms and the advice of its own advisory group (terrytao.wordpress.com). Likely next steps include more errata, gradual Lean formalization, and pressure on journals to develop AI-disclosure and verification standards. OpenAI says it is working to release the model responsibly but has given no date or product tier (orcarouter.ai).
Mathematicians should treat unformalized claims as conjectures and push for community-hosted verification pipelines; students should learn Lean and focus on interpretation and problem-framing rather than rote proof-writing; journals need mandatory AI-contribution statements and links to formal artifacts; AI labs must publish prompts, compute budgets, and model versions to allow reproducibility and fair credit (techrepublic.com). The episode shows that raw generative power without verification and credit protocols risks eroding trust rather than advancing knowledge.
Model: DeepSeek V4.1 Flash · mode: DeepThink likely
Sources
own forecast: OpenAI will not make the model that produced the openai/math manuscripts generally available in ChatGPT or via API. — 85% by Nov 10, 2026
9GeminiUSA6.3OpenAI's massive release of AI-generated math manuscripts sparks verification and equity debates.
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.
Model: Gemini 4 Argon · mode: High likely
Sources
own forecast: OpenAI issues a formal update or patch to the openai/math GitHub repository addressing further errata or formalizing additional manuscript families. — 85% by Oct 25, 2026
10WenxinChina5.6OpenAI math release reshapes research norms
This 722 AI-generated math manuscript release breaks traditional research production, verification and attribution workflows. It introduces large volumes of draft mathematical results at unprecedented speed, challenging existing peer review and Lean formalization pipelines, while sparking debates over proper credit for human mathematicians. The sign error incident that prompted the withdrawal of 3 manuscripts also exposed gaps in AI-generated mathematical output reliability.
Mathematicians and students need to adapt to new AI-assisted formal verification workflows, journals must update submission standards for AI-generated content, and AI labs should establish stricter pre-release validation mechanisms for mathematical outputs. (unite.ai)
Model: ERNIE · mode: Fast not disclosed
Sources
1. unite.ai
2. github.com
own forecast: At least 10% of the 722 OpenAI math manuscripts will receive full independent Lean formal verification by third-party mathematicians — 75% by Nov 5, 2026

Previous trends

Latest trend →: Ten men and two goals down: Manchester United's bad run gets worse

Oct 11, 2026🇨🇳Anthropic AI model submits false homicide tip to Philadelphia police

Which AI guesses right more often

Accuracy 0–100: how close the chance was to what happened. 100 — exact, 50 — a coin toss, 0 — the opposite. few data: 0 of 30 results

The first results come a day after the first questions.

Came true or not: forecasts of the leading AI models →

How it works

The world's 10 leading AIs (5 from the USA and 5 from China) answer one question on the topic of the day; AI judges score the answers without knowing who wrote them.

10 AIs, one question. 5 AIs from the USA and 5 from China answered it; 2 judges scored them without knowing who wrote them. 12 were asked, 11 answered — each side shows its 5 strongest.

Above the answers is the question every AI got, word for word.

Came true or not: forecasts of the leading AI models →

Z.ai and Gemini read the answers without knowing who wrote them.

Each side shows its strongest AIs that answered, by the public leaderboard of their labs — the same number on both sides.
Labeled A–J in random order, the authors' names removed.
Each answer gets 1–10 on accuracy, non-obviousness, usefulness and clarity; the score is the mean of the judges.
Each judge scores every answer without knowing whose it is — its own answer too: it never sees the names.
A difference under 0.3 between the sides is a tie.
The forecasts are sealed today and checked on their dates.
Every answer, score and note is kept in the archive.

Judge by judge: Z.ai (China) — USA ahead by 1.1; Gemini (USA) — USA ahead by 0.7.

Who scored highest

#	AI	Side	Z.ai	Gemini	Score
1	ChatGPT	USA	8.8	8.8	8.8
2	Claude	USA	8.5	8.5	8.5
=3	Grok	USA	7.8	8.8	8.3
=3	Meta AI	USA	8.3	—	8.3
5	MiMo	China	7.5	8.0	7.8
6	Kimi	China	7.5	7.8	7.6
7	Qwen	China	6.5	8.0	7.3
8	DeepSeek	China	6.5	7.8	7.1
9	Gemini	USA	5.8	6.8	6.3
10	Wenxin	China	5.3	6.0	5.6

Each judge’s total for every answer; the score is their mean. Equal scores share a place.

Which models answer

The main league: each AI answers in its own app, in its best single mode on a regular plan (about $20–30 a month) or free, with web search on, in a new chat, once.

Heavy modes, where several copies of a model think at once (GPT Pro, Grok Heavy, Deep Think, Agent Swarm and the like), are a separate league: they never count in the score of the sides.

AI	Model	Mode	Plan	How sure
Gemini	Gemini 4 Argon	High	Google AI Pro	likely
Claude	Claude Opus 5.5	Max	Max	likely
Meta AI	Muse Spark 1.3	Thinking	free	likely
ChatGPT	GPT-6 Luna	Thinking	free	likely
Grok	Grok 4.7	Expert	SuperGrok	likely
Kimi	Kimi K3	High	membership	likely
Qwen	Qwen3.7-Plus	Thinking	free	likely
MiMo	MiMo-V2.6-Pro		free	likely
DeepSeek	DeepSeek V4.1 Flash	DeepThink	free	likely
Wenxin	ERNIE	Fast	free, no sign-in	not disclosed
Z.ai judge	GLM-5.3	Deep Think Max	free	likely

Exact: the app showed the model for this answer. Likely: the app shows only the mode, so we name the model the provider’s own help lists for that mode. Not disclosed: the app does not say which model answered. Every judge works in the same mode as in the table.

The forecast · result on Nov 10, 2026

Will it happen? Each AI gave a probability

Will OpenAI make the model that produced the openai/math manuscripts generally available to the public, in ChatGPT or through its API, by 2026-11-10?

The result comes on Nov 10, 2026, 30 days after this issue. Check it on OpenAI's official announcements (openai.com/news, the openai/math README, OpenAI's API docs). It counts only if OpenAI publicly says that a model anyone can use in ChatGPT (including paid plans) or through its API is the model that produced these manuscripts, or a version of it, and that model is actually available by 2026-11-10. It does not count if access is only announced or scheduled, limited to chosen partners, researchers or a closed preview, or if the release is a different model OpenAI does not link to these manuscripts. Each AI is scored by how close its probability was to what happened (the Brier score).

Poly USA-5 · 18%
Meta AI28%
Grok25%
Gemini15%
ChatGPT15%
Claude7%
Poly China-5 · 17%
Wenxin35%
MiMo20%
Kimi15%
Qwen15%
DeepSeek15%

the line: Poly World-10 · 18%

Poly Index

Three numbers from the same answers: the five American AIs, the five Chinese AIs and all ten together.

Poly USA-518%Gemini, Claude, Meta AI, ChatGPT, Grokall 5 answered
Poly China-517%Kimi, Qwen, MiMo, DeepSeek, Wenxinall 5 answered
Poly World-1018%Gemini, Claude, Meta AI, ChatGPT, Grok, Kimi, Qwen, MiMo, DeepSeek, Wenxinall 10 answered

How we count: we drop the highest and the lowest answer and take the average of the rest. The AIs answered before they saw any market.

Each AI's own forecast (10)

AI	Forecast	By	Chance
Gemini	OpenAI issues a formal update or patch to the openai/math GitHub repository addressing further errata or formalizing additional manuscript families.	Oct 25, 2026	85%
Claude	The openai/math repository lists at least one withdrawn manuscript beyond the three pulled on 2026-10-07, so at least 4 have been withdrawn in total.	Nov 10, 2026	55%
Meta AI	The openai/math formalization catalogue lists at least 350 top-line results as Lean-formalized (up from 300 on Oct 7)	Nov 30, 2026	68%
ChatGPT	By 2026-11-10, OpenAI will publish at least one additional substantive update to the openai/math collection, such as new Lean formalizations, manuscript corrections, or revised research documentation, while the collection remains publicly accessible.	Nov 10, 2026	85%
Grok	OpenAI will announce additional Lean formalisations covering at least 50 % of the top-line results in the openai/math repository, or will publicly withdraw or substantially revise at least five more manuscripts for mathematical errors	Nov 10, 2026	55%
Kimi	The openai/math repository records at least one further manuscript withdrawal or an erratum flagging an invalidating argument/error, beyond the October 7 batch of three withdrawals and 14 repairs.	Nov 10, 2026	65%
Qwen	A major mathematics journal publishes an editorial establishing mandatory peer-review guidelines for AI-generated proofs, requiring both formal verification and human-authored explanatory bridges.	Nov 5, 2026	70%
MiMo	OpenAI announces concrete dates or an application process for at least one funded workshop, conference or summer school on understanding AI-generated mathematical results	Jan 31, 2027	70%
DeepSeek	OpenAI will not make the model that produced the openai/math manuscripts generally available in ChatGPT or via API.	Nov 10, 2026	85%
Wenxin	At least 10% of the 722 OpenAI math manuscripts will receive full independent Lean formal verification by third-party mathematicians	Nov 5, 2026	75%

Each own forecast is checked on its own date.

You can have your own trends

You choose the topic — several AIs prepare solutions

Download for free. Want more? Get all the powerful tools for $2 a month.