OpenAI AI Math Proofs Face Lean Verification Backlash

What happened

OpenAI's push to show that its models can do research-level mathematics has hit a wall of skepticism from the people best placed to judge it: working mathematicians. Over the past few months the company has released a growing volume of AI-generated results, ranging from a handful of headline problems in August to hundreds of manuscripts by October. Critics now say the volume has far outpaced the verification.

The sharpest number comes from tallies of how many claims have been formally checked in Lean, the proof assistant that can mechanically confirm a mathematical argument. Only about 42% of the claims in a 377-result release reportedly pass Lean checks. The rest sit in an uncomfortable middle ground: published and attributed to an AI system, but neither machine-verified nor refereed the way journal work normally is 1. Terence Tao and Tristan Buckmaster are among those who have raised concerns about both verification and attribution 1.

The scale figures vary depending on which release is being counted. Separate reporting describes a Tao-led advisory group examining 722 OpenAI manuscripts 3. The two numbers likely reflect different batches or stages rather than a contradiction. Either way, the volume is far beyond what conventional peer review can absorb quickly.

The Navier-Stokes flashpoint

The most ambitious claim concerns Navier-Stokes existence and smoothness, one of the Clay Mathematics Institute's Millennium Prize problems. OpenAI reportedly produced it using roughly 10,000 AI agents over about 88 hours 3. The Clay Institute had not independently verified or accepted the claim when the BBC reported on it 3.

The problems go beyond slow review. Researchers from the University of Cambridge and King's College London found discrepancies between OpenAI's written explanation of the Navier-Stokes work and the code used for its computer verification, according to TechCrunch reporting from October 8 4. In addition, three separate research drafts were retracted over calculation errors 4.

A timeline of escalating claims

The trajectory helps explain the friction. In August, OpenAI's claims covered 10 major problems, including work on non-sofic groups and Erdős problems 146, 180 and 183, none of them independently confirmed 3. Model training reportedly began in late August. By September, the company said more than 100 problems had been solved or advanced, which prompted the formation of the advisory group 3. The Navier-Stokes claim followed, and by October the manuscript count had reached 722 3.

The pace of announcements has consistently run ahead of external confirmation.

Why a green checkmark isn't enough

One might assume formal verification settles the matter. A recent study complicates that assumption. It finds that AI can produce Lean-verified proofs that do not faithfully reflect the argument presented in the human-readable paper. The result is what the author calls a dangerous illusion of certainty 2. A proof assistant confirms that some formal statement holds. It does not confirm that the formal statement matches the theorem being claimed, or that the prose explanation describes what the code actually does 2.

That is exactly the gap the Cambridge and King's College researchers flagged in the Navier-Stokes case, where the explanation and the verification code diverged 4. So the 42% pass rate is a floor of formal confidence, not a ceiling of correctness. Even passing results need human scrutiny of what was actually formalized.

Readability compounds the problem. Many of the proofs are reportedly hard for mathematicians to follow as written. Researchers have asked OpenAI for more access to the model, its prompts, and its intermediate reasoning, and those requests had not been fully answered as of recent reporting 1.

A divided community

The criticism is not uniform. A mathematics group identified as AGMAI reportedly objected to the whole practice of testing proprietary frontier models against major open problems. It argued that research mathematics should not be reduced to checking whatever a closed lab chooses to publish on its own timetable 1. That is a more fundamental objection than Tao's group raises. The advisory group is engaging with the manuscripts rather than rejecting the exercise 3.

OpenAI is also not alone. AI-assisted mathematics expanded across the field in 2026, with hundreds of AI-involved submissions on arXiv and work from Google DeepMind and Meta. Meta released a preprint titled "Learning to Discover Interesting Mathematics" on September 23 5. The verification and attribution questions raised by OpenAI's releases will apply to everyone.

The reading

The underlying capability may be real. Some results will likely survive scrutiny. But OpenAI has inverted mathematics' normal order: claim first, verify later, at a volume that overwhelms reviewers. Retractions, code-explanation mismatches, and a sub-majority formal pass rate suggest the announcements have outrun the evidence.

The credible path forward is unglamorous. It means releasing prompts and intermediate steps, auditing that formal statements match the claimed theorems, and letting bodies like the Clay Institute rule before declaring victory. Until then, "solved" should be read as "submitted."