Estimated reading time: 6 minutes · Last updated:
Dartmouth mathematicians and students are debating what to make of OpenAI’s Oct. 6 release of 722 AI-generated math manuscripts organized into 372 groups. The surge includes claims touching several of the seven Millennium Prize Problems, each carrying a $1 million award, and has forced researchers to separate two questions: are the results correct, and if so, do they advance mathematical understanding? As first reported by The Dartmouth, faculty said many manuscripts are currently unreadable, only a subset has been formalized in the proof assistant Lean, and several high-profile items were later withdrawn.
Just answering the questions is not the point.
Erik van Erp, Dartmouth mathematics professor
Key takeaways
- Scale of the release: OpenAI published 722 manuscripts on Oct. 6, organized into 372 groups covering fields such as number theory, geometry and theoretical computer science.
- Verification status: The repository contains 719–722 reported findings and, as of the reporting, 300 of the findings have been formalized in Lean.
- Withdrawals and corrections: OpenAI withdrew three manuscripts related to the Hodge conjecture on Oct. 7 and listed revisions for 14 others.
- Faculty concerns: Dartmouth professors including Erik van Erp and Asher Auel described many papers as 'impenetrable' or 'almost completely unreadable' and questioned whether new insight had been produced.
Table of contents
- Key takeaways
- What OpenAI released and how the community has begun to check it
- Why many Dartmouth mathematicians are sceptical
- Students and early-career researchers weigh opportunity and risk
- Next steps: verification, organization and academic response
- How this could play out
- What to be careful about
- Frequently asked questions
What OpenAI released and how the community has begun to check it
On Oct. 6, OpenAI posted a large collection of machine-generated manuscripts to GitHub. The release spans disparate areas of mathematics and claims progress on problems that include some of the seven Millennium Prize Problems, established in 2000 with a $1 million prize for a correct solution.
The items were grouped into 372 sets of related findings; the publicly viewable repository lists slightly different counts in places, with 719 and 722 figures appearing in the same record. Community scrutiny has focused on formal verification: mathematicians enter proofs into the proof assistant Lean to break arguments into mechanically checkable steps. According to the reporting, 300 of the repository’s findings have been formalized in Lean so far.
Formalization converts informal arguments into a sequence of definitions and lemmas the software can verify. That step strengthens confidence but does not replace human judgement about conceptual novelty or correctness beyond the formalized portion.
Why many Dartmouth mathematicians are sceptical
Several Dartmouth faculty interviewed by the reporter said the manuscripts are often unreadable and do not yet supply the explanatory route that makes a proof valuable to researchers. Mathematics professor Erik van Erp said the value of a result rests not only on a correct answer but on what is learned in the process, and he described some papers’ introductions as impenetrable.
Asher Auel told the reporter that one high-profile manuscript claiming to disprove a conjecture in algebraic geometry looked poised to overturn assumptions if true, but he found the work "almost completely unreadable" and said it is impossible to tell whether a viable strategy is being offered. Those judgements underline a practical gap: machine output can state claims and lines of reasoning, but human experts are needed to parse context, spot subtle errors and determine whether new techniques are present.
Dartmouth postdoctoral researcher Yao Xiao offered a different register: she sees AI as accelerating background work and surfacing connections that enable human researchers to contribute meaningfully. That division — some viewing AI as a tool, others as an unreadable black box — is the present fault line.
Students and early-career researchers weigh opportunity and risk
Undergraduate researcher Benjamin Singer said his first reaction was scepticism and that he learned of the release through a student group chat. Singer described the repository as containing many interesting ideas but criticized OpenAI’s public response after the withdrawal of three Hodge-related manuscripts on Oct. 7, calling the company’s explanation—describing a "sign error"—an underplay of the issues.
Singer said AI could make mathematical exploration accessible to people without formal training and that he has received AI-generated work from external readers that appears correct. Still, he worries that senior researchers’ reactions will shape the prospects of younger mathematicians, and he noted anxiety among students about finding a place in a changing field.
Former college president and professor Phil Hanlon framed the discussion in human terms, arguing that discovery carries an emotional thrill for researchers and collaborators that would be diminished if machines did the end-to-end work. Those concerns sit alongside pragmatic ones about careers and evaluation.
Next steps: verification, organization and academic response
The immediate task for the community is triage: identify manuscripts that deserve deeper study, formalize proofs where feasible and surface any novel ideas. The reporter notes that only a minority of the findings are formalized in Lean, and several manuscripts were revised or withdrawn soon after release.
Beyond verification, mathematicians called for better organization of the flood of claims so the field can separate genuine advances from errors or recycled material. That includes journal review, community-led reading seminars and improved indexing so specialists can find work relevant to their subfield.
Open questions remain about the unreleased OpenAI model that produced the manuscripts and whether its outputs can be harnessed as a reliable research assistant. For now, Dartmouth interviewees agreed that human expertise remains essential both to verify correctness and to extract the insights that make a result useful.
| Item | Count / example | Community status |
|---|---|---|
| Manuscripts released | 722 (Oct. 6) | Publicly posted on GitHub |
| Grouped findings | 372 groups | Organized by related topics |
| Formalized in Lean | 300 findings | Mechanically checked; minority of total |
| Withdrawn manuscripts | 3 (Oct. 7) | Withdrawn by OpenAI; revisions listed for 14 others |
How this could play out
The case for
- AI tools accelerate background checks and idea generation, making some research workflows faster, as Dartmouth postdoctoral researcher Yao Xiao noted.
- Formal verification in Lean could validate key results and provide a searchable corpus of machine-checked arguments for future work.
- Organized review by journals and community groups could turn the release into a set of usable leads rather than noise.
The case against
- If many manuscripts are unreadable or incorrect, the release could waste researchers’ time and generate confusion about priority and credit.
- High-profile errors or inadequate disclosure from a single organisation could damage trust between AI developers and the mathematical community.
- Early-career researchers may face uncertain evaluation criteria if machine-generated claims complicate hiring and publication norms.
What to be careful about
- Misplaced trust in unverified machine output that has not been formalized in Lean.
- Reputational harm to researchers if AI-produced claims are mistaken for human-authored original work without clear attribution.
- Career disruption for junior researchers if institutions do not settle how to assess AI-assisted results.
The bottom line
The OpenAI release is a stress test for mathematical practice: it shows how rapidly generative systems can produce claims, and it exposes the community’s dependence on human judgement, explanation and formal verification. Dartmouth interviewees agree on one point: correctness alone is not sufficient. For a result to shape a field it must be checkable, teach new techniques and be integrated into existing literature — tasks that, for now, still require mathematicians’ expertise.
What to watch
- Watch for which additional manuscripts are formalized in the proof assistant Lean; no date has been set.
- Watch for journal editors or professional societies to publish formal reviews or errata addressing OpenAI’s submissions; no date has been set.
- Watch for any public technical response from OpenAI about the withdrawn Hodge-conjecture manuscripts beyond the note dated Oct. 7; no date has been set.
Frequently asked questions
How many manuscripts did OpenAI release and how were they organised?
OpenAI released 722 manuscripts on Oct. 6 and grouped them into 372 sets of related findings; the repository shows slightly differing counts in places, with 719 appearing alongside 722.
What does it mean that 300 findings were formalized in Lean?
Formalization in Lean means proofs were rewritten as mechanically checkable steps in the proof assistant Lean; the reporting says 300 of the repository’s findings reached that stage, which increases confidence but does not by itself guarantee conceptual novelty.
Did OpenAI withdraw any manuscripts?
Yes; OpenAI withdrew three manuscripts related to the Hodge conjecture on Oct. 7 and listed revisions for 14 others in its public history.
Related reading
