OpenAI has released nearly 400 AI-generated mathematical results spread across more than 700 manuscripts, covering everything from combinatorics, geometry, and number theory to topology and mathematical physics. The sheer volume is so large that it calls for a guide to navigating the GitHub repository.
The problem is that verification remains incomplete. Of the 719 manuscripts, only 300 results have been formalized in Lean, or roughly 42%. Even when code is available for inspection, researchers still need to check whether the proof actually matches the claims made in the paper. Some mathematicians therefore believe it could take years just to fully understand everything.
What OpenAI Released This Time
OpenAI released nearly 400 results from AI, distributed across 719 manuscripts and covering fields ranging from combinatorics, geometry, number theory, and algebra to topology, probability, and mathematical physics.
The GitHub repository is so large and diverse that reading through it all at once is difficult. OpenAI therefore published a guide to help researchers navigate the collection and select work from each field more easily.
When Reading the Table of Contents Takes Almost an Hour
Imagine needing almost an hour just to read around 40 pages of tables of contents and abstracts. That is what many mathematicians faced after OpenAI released this large body of work all at once.
The problem is not simply that there is a lot of material. It is also the challenge of distinguishing which work is credible, which requires further verification, and which may be “AI slop.” As a result, the initial assessment itself becomes almost as substantial as reading actual research papers.
OpenAI is moving from using AI to write mathematics toward releasing work that affects the entire field. More than 30 researchers therefore do not see this as merely a demo, but as an event that will take a long time to fully understand.
Responses such as “Staggering,” “Overwhelming,” “Unprecedented,” “Surreal,” and “Pure insanity” all reflect the same picture: the scale of this release is so large that the field must rethink how it screens, verifies, and collaborates with AI.
How This Differs from OpenAI’s Previous Mathematical Work
Earlier work was criticized for poor or nearly nonexistent citations. The latest release is much larger, requiring extensive time for screening and verification.
| Factor | OpenAI’s previous work | The latest release |
|---|---|---|
| Number of works | Not specified | Nearly 400 results across 719 manuscripts |
| Scope of content | Not clearly specified | Multiple fields, such as geometry, number theory, and topology |
| Citations | Criticized as poor or nearly nonexistent | Still needs to be checked individually |
| Formalization | Lean used to check some claims | 300 of 719 manuscripts formalized |
| Problems found | Difficult to read and lacking clarity about credit | Verification levels vary, and some Lean formalizations do not match the claims in the manuscripts |
From Lean to Verification That Still Requires Humans
Lean functions like a checkpoint that helps confirm whether a proof works computationally. However, people still need to check whether the code proves the same thing claimed in the manuscript.
Imagine a researcher opening the GitHub repository and having to screen a large number of manuscripts. Having 300 of the 719 manuscripts already formalized makes it possible to begin checking more quickly, but it still does not cover the entire collection.
The problem is that the level of verification varies. Some claims have computer-verifiable proofs, but the content may not match the claims in the manuscripts. Researchers must therefore continue reading, checking consistency, and evaluating quality themselves.
The Competitor to New Knowledge May Not Be Another AI
The issue is not how quickly OpenAI, ChatGPT, or Claude can generate work. It is the verification burden that follows: difficult-to-read content, limited understanding, and potentially inadequate attribution of researchers.
| Factor | OpenAI’s research collection | Using ChatGPT | Using Claude |
|---|---|---|---|
| Overall risk | The volume makes initial evaluation difficult | Increase in low-quality AI content | Increase in low-quality AI content |
| Reading and understanding | Difficult to read and requires time to verify | Confusing and demonstrates limited understanding of the content | Confusing and demonstrates limited understanding of the content |
| Attribution | Previously criticized for citing researchers | May inadequately credit researchers | May inadequately credit researchers |
What Is Exciting, and What Remains Concerning
OpenAI has released mathematical work spanning multiple fields, with some of it formalized in Lean. This allows mathematicians to verify its correctness computationally. The collection may therefore open the door to interesting discoveries, but it still requires careful examination.
Pros
- +Covers multiple fields of mathematics
- +Lean formalization helps verify the logic
- +May contain new results worthy of study
Cons
- −The verification burden falls on mathematicians
- −Quality and levels of verification are inconsistent
- −Risk of error-prone and difficult-to-check AI slop
The Cost That Does Not Appear on the GitHub Repository
The real cost is mathematicians’ time, not storage space. Simply going through around 40 pages of tables of contents and abstracts already takes almost an hour. After that, researchers still need to read the work, check whether Lean proves what is claimed, and separate genuine results from errors.
It may take years to fully understand this collection. If OpenAI releases another batch before the previous one has been completely checked, the burden will only accumulate. Researchers must choose between reading content that may be AI slop, waiting for others to verify it, or waiting for more formalization to become available.
Who Should Follow This Research Collection, and Who Should Wait
This collection is suitable for mathematicians and researchers who want to explore work across multiple fields and verify claims with Lean themselves. However, they should be prepared for each piece to have a different level of verification.
Those looking for results they can immediately read and trust should wait. Even when formalization is available, researchers must still check whether the proof matches the claim.
Made for
- Mathematicians who want to explore work across multiple fields
- Researchers prepared to verify results with Lean
Think twice
- People who want to read work that has already been checked quickly
Skip this one
- General readers who do not want to sift through a large volume of work — wait for summaries from researchers
The Important Question May Not Be How Many Problems AI Can Solve
The value of this release may lie not in the number of results, but in the mathematical community’s ability to verify and understand them. Many results are still in the process of being confirmed, and even with Lean, researchers must check whether the proof actually matches the claims in the manuscripts.
The true impact may therefore be measured by how effectively researchers can screen the work, connect results across fields, and separate credible research from AI slop. This is not merely about seeing what AI can do now, but about testing how the field will cope with this level of speed and volume.