Less than 24 hours after OpenAI published 722 AI-generated math manuscripts on October 6, 2026 — a release we broke down in a separate post — the story took a sharp turn. On October 7, OpenAI disclosed that a sign error in a single supporting construction had quietly undermined three of those manuscripts, all connected to its Hodge conjecture results, forcing their withdrawal along with 14 lesser revisions elsewhere in the batch. By the end of the same week, a Fields Medalist was warning of 'proof indigestion,' an organization representing mathematicians was calling for an outright boycott, and more than 800 researchers had put their names to the statement, accusing OpenAI of rushing unvalidated work into the world (Sources: Retraction Watch, The Decoder).
None of this is really a story about one bad proof. It's a story about what happens when a single company can generate more specialist mathematics in a day than the global mathematics community can realistically check in months — and whether that's a gift to the field or a problem for it.
Quick summary: On October 7, 2026 — one day after OpenAI's 722-manuscript math release — the company withdrew 3 manuscripts related to the Hodge conjecture over a sign error (a +1 where a -1 belonged, in a geometric construction) and issued 14 other revisions. The same week, the Association for Human Mathematics, which press coverage describes as chaired by Fields Medalist Terence Tao (a title not confirmed on the organization's own site), and whose statement Tao posted as a guest contribution on his own blog, called for mathematicians to stop working with OpenAI, describing the release as 'a demonstration of power, not scholarship.' The Association's own statement gathered more than 800 signatures within days, with signatories accusing OpenAI of rushing unvalidated work past peer review. Tao also published a personal essay warning of 'proof indigestion' and arguing for what he calls a 'Math 2.0' era, where autonomously-generated proofs change what the discipline should actually value.
A +1 where there should have been a -1
The technical failure itself was narrow: reporting describes it as a sign error — an incorrect +1 instead of -1 — inside a geometric operation supporting the model's Hodge conjecture results. That single flipped sign was enough to break the logical chain underneath two Hodge-conjecture manuscripts and a third, related one, all three withdrawn on October 7. OpenAI additionally issued 14 revisions across other parts of the 722-manuscript batch — smaller fixes that didn't require full withdrawal but did require correction, all surfacing within a single day of outside and internal scrutiny (Sources: Retraction Watch, implicator.ai).
Three retractions out of 722 manuscripts is, numerically, a small fraction. But the speed of the discovery — under a day — is exactly what's fueling the louder argument: if a sign error buried in a single geometric construction surfaced within hours of a few people actually looking closely, what's sitting unexamined in the other 719 manuscripts, several of which touch subject areas where only a handful of specialists worldwide are qualified to check the work line by line?
The boycott call
The Association for Human Mathematics — a group founded in September 2026 by 25 Fields Medalists, and described in press coverage (including The Decoder) as chaired by Terence Tao, though AHM's own site names no formal chair — issued a statement in the days following the release urging mathematicians to stop collaborating with OpenAI, framing the simultaneous publication of 722 manuscripts as 'a demonstration of power, not scholarship.' Tao posted the statement to his own blog as an explicit guest post from the Association, signaling the mathematics establishment's response wasn't just a fringe reaction (Source: 36kr).
That statement itself gathered more than 800 signatures within days, with signatories accusing OpenAI of rushing unvalidated work past the normal checks the field relies on — peer review, specialist scrutiny, slow replication — in favor of a single, attention-grabbing mass release (Source: The Decoder).
Tao's 'Math 2.0' argument
Tao's own commentary — a "Math 2.0" talk he gave at Caltech on October 9, written up on his own blog the next day — goes further than process complaints. He warned of what he called 'proof indigestion' — a field being handed more purported results than it has the specialist capacity to digest — and argued that problems are increasingly being solved by AI prompters with no interest in the broader field once their target is solved, a pointed distinction between generating a correct-looking proof and actually contributing to how mathematicians understand a problem. His proposed response is a shift he's calling 'Math 2.0': an era in which solving individual open problems stops being the discipline's central measure of progress, precisely because a sufficiently capable model can now generate solutions faster than humans can verify, contextualize, or build on them (Source: kingy.ai).
It's worth separating two different objections here, because they get conflated in headlines. One is about correctness — did OpenAI ship broken proofs? (Answer: a small number, yes, and it corrected them fast once caught.) The other is about process and incentives — should a lab be allowed to dump 722 unvetted specialist manuscripts into a field at once, regardless of whether most of them turn out correct? That second objection is really what Tao and the Association for Human Mathematics are pressing on, and it doesn't go away even if every remaining manuscript eventually checks out.
Why this is bigger than mathematics
The underlying tension generalizes well beyond math departments. As AI models get better at generating plausible-looking, technically dense output in any specialist field — law, medicine, structural engineering, materials science — the same question recurs: who verifies it, how fast, and what happens in the gap between publication and verification? Mathematics is arguably the best possible field to stress-test this in, because proofs are at least formally checkable (OpenAI leaned on this itself, including Lean-formalized verifications for many of its 722 manuscripts) — most other fields don't have anything close to that level of built-in auditability.
- Mathematicians and researchers deciding whether to engage with, cite, or build on any of OpenAI's released results.
- Anyone in a specialist field watching mathematics as a bellwether for how AI-generated expert output will be handled elsewhere.
- Journal editors and conference organizers thinking through policy on AI-assisted or AI-generated submissions.
- General readers trying to calibrate how much to trust 'AI solved X' headlines going forward — the honest answer, per this episode, is: check who's verified it, and how recently.
01What exactly caused OpenAI to withdraw three of its math manuscripts?
A sign error — an incorrect +1 instead of -1 — inside a geometric construction supporting its Hodge conjecture results. The error invalidated two Hodge-conjecture manuscripts and a third related one; OpenAI withdrew all three on October 7, one day after the original release, and issued 14 further revisions elsewhere in the batch.
02Who is calling for a boycott of OpenAI, and why?
The Association for Human Mathematics (AHM), founded in September 2026 by 25 Fields Medalists and widely reported as chaired by Terence Tao, called on mathematicians to stop working with OpenAI, describing the simultaneous release of 722 manuscripts as 'a demonstration of power, not scholarship' rather than a genuine scholarly contribution. The statement gathered more than 800 signatures within days.
03What is Terence Tao's 'Math 2.0' idea?
Tao argues that as AI models can generate purported proofs faster than the mathematics community can verify them, the field needs to stop treating 'solving an open problem' as its central measure of progress, and instead prioritize verification, context, and understanding over raw proof production.
04Does this mean OpenAI's math results are wrong?
Not wholesale — only 3 of 722 manuscripts have been withdrawn so far, plus 14 revisions. The bigger unresolved question isn't whether most of the release is correct, it's whether the field can verify 722 specialist manuscripts fast enough for the release to be trustworthy at the moment it's published, rather than months or years later.
05Is this OpenAI's first controversy over AI-generated research?
It's the most prominent one in mathematics specifically. OpenAI's earlier, smaller claim — disproving a 1946 Erdős conjecture in May 2026 — did not trigger a comparable reaction, likely because of its much smaller scale.
Stories like this are a good reminder that even the most capable AI model benefits from being checked against another one — a second, independent opinion is basically free when you're not locked into a single provider. LumiChats puts 40+ premium AI models, including the latest from OpenAI, Anthropic, Google, and others, in one place, so you can cross-check a claim, a proof, or a piece of research-heavy output across multiple models before deciding how much to trust it.
