Listen to this article
AI may have been used to generate this audio.
Press Space or K to play or pause, and J or L to go back or forward 15 seconds.Sources checked 8 October 2026.
On 6 October, OpenAI published a large collection of mathematical manuscripts from an internal model. The repository describes results at different stages of verification and warns that some unformalised results could have issues. That qualification belongs beside any headline about the release. [1][2]
A manuscript is a claim someone can inspect. Publishing hundreds of them gives the mathematical community a considerable amount of inspection work. A reader deciding whether a particular result is useful needs more than an impressive total.
The catalogue separates manuscripts from families of related results and reports partial formalisation. Its figures describe the repository, rather than an independent audit. [2] A reader still has to follow the particular result they want to use.
A headline that turns every file into a separate solved problem has already changed the meaning of the source. So has a headline that treats every unformalised argument as false. Mathematicians have long checked written arguments without proof assistants. The relevant question for any one result is which checks it has received and what remains uncertain.
The Association for Human Mathematics objected to the release and called on mathematicians to stop working with OpenAI. Its statement concerns the publication model and the community’s priorities. It represents that association’s position, rather than a verdict on the correctness of every manuscript. [3]
What reviewing a result involves
Imagine that one paper appears relevant to a problem you have been working on. Before using it, you need to know exactly what it establishes. A theorem with an extra assumption may be useful while falling short of the result described in a social-media post. The distinction can disappear in a sentence announcing that a problem has been solved.
Then there is the relationship to earlier work. A proof might be new, a reformulation of an existing argument, or dependent on a result that should have been cited. These questions require attention from people familiar with the subject. Downloading the PDF does not resolve them.
A reviewer may do several of these checks while reading the same argument. Finding a hidden assumption can change the theorem statement, the search for earlier work and the explanation offered to other researchers. Review needs a record of those decisions, without pretending that every result must pass through the same sequence of committees.
There is another distinction that gets lost in the public argument. A correct theorem can still have a mistaken claim of novelty. If an older paper proves the same result, the new argument may offer a simpler route or a useful method, but that contribution needs a different description. Formal checking cannot search the historical literature and decide who should receive credit.
Conversely, a correction need not erase the value of an entire paper. It could narrow a hypothesis, repair a proof step or change which earlier theorem is needed. Readers need a visible account of the change to decide which uses remain sound. Treating every revision as either vindication or collapse makes that ordinary scholarly work harder to follow.
For a researcher trying to build on the release, I would begin with a single dependency. If my argument needs theorem T from version v, I need to record that dependency and the assumptions I imported with it. If T changes, I can identify what needs another review. A link to the collection’s front page would not give a future reader enough information to reproduce that decision.
Who takes responsibility for that work
The Advisory Group on Mathematics and Artificial Intelligence recommends that labs support the human work needed to understand substantial AI output. It also says that the mathematical community should lead that understanding. OpenAI says it will fund workshops, conferences and programmes around major AI-produced results. [4][1]
That leaves practical choices. A researcher needs to know which version they reviewed, whether a later correction affects it, and how to report a problem. Someone coordinating review needs to avoid having several people repeat the same check while another consequential claim receives none.
A useful record would begin with the exact theorem and manuscript version. It would name the available proof artifacts, record what a reviewer checked and keep unresolved questions visible. A broad label such as “verified” would be too coarse if it hid whether the check concerned a formal proof, a transcription or the headline claim.
Support also has to match the work being requested. A conference can gather the right people, but a difficult formalisation may require sustained engineering and a careful comparison with the written theorem. An explanatory paper may need a different set of readers. Paying for an event and paying for the months of work around it are separate commitments.
There is a reasonable objection to demanding that every result wait for complete human understanding before release. Other researchers may be able to use or explain it sooner if they can inspect it. The answer I favour is explicit status, preserved versions and funded follow-up. That lets people investigate while keeping a provisional result from quietly acquiring the status of a settled one.
The public response should also leave room for mixed assessments. A mathematician can consider a result significant and dislike the publication process. Another can welcome the release and identify a serious gap in one proof. A useful account of this episode will need those result-level judgements, including the ones that refuse to fit an argument for or against AI.
The incentive problem is straightforward. The organisation announcing a large release receives immediate attention. The person finding a missing citation or explaining a difficult argument may spend days on a small portion of it. If the release depends on that work, funding and credit should make room for it.
A review that identifies a missing assumption, supplies the right citation or makes a method understandable should be recorded alongside the result it improves. Readers need to see what changed and who did that work.
Sources
One email when we publish. Research, product decisions, and what teams report back.
