Some OpenAI Math Discoveries Are 'Jew els,' if Verified. But Mathematicians Are Wary of 'Slop

Some OpenAI Math Discoveries Are 'Jew els,' if Verified. But Mathematicians Are Wary of 'Slop.'

**OpenAI says it solved more than 370 math problems. Mathematicians say there could be breakthroughs in the bunch, if verified — but that's no easy task with unpolished work.**

-

--

## The Night the Proofs Arrived

Let me tell you about a professor named Andrew. He teaches mathematics at a university in the Midwest. He's spent his career working on problems that most people can't pronounce and even fewer can solve. He's used to the slow grind of research — the years of false starts, the small breakthroughs, the occasional moments of genuine discovery.

Last week, he opened his laptop and found **722 manuscripts** sitting on GitHub. All generated by an AI model he couldn't access. All claiming to solve or advance some of the hardest problems in his field.

His first reaction was curiosity. His second was dread.

"I don't have time to read all of this," he told me. "And even if I did, I'm not sure I'd understand it. The AI writes proofs the way a machine writes poetry — technically correct, structurally sound, but missing something essential about *why* it's true."

That's the dilemma facing mathematicians across America right now. OpenAI has released what it calls a treasure trove of mathematical discoveries. Some of them, according to the experts, may be genuine breakthroughs — "jewels" in the rough. But the sheer volume, the lack of transparency, and the unpolished nature of the work have created a crisis of verification that the field is struggling to handle.

---

## What OpenAI Actually Released

Let's get the facts straight, because the numbers are staggering.

On October 6, 2026, OpenAI published **722 mathematical manuscripts** on GitHub, grouped into **372 "result families"** covering topics from algebra to theoretical computer science to geometry .

The model that produced these results was an **internal, unreleased system**. OpenAI says it attempted roughly **4,000 problems** and kept the outputs it judged significant .

The average result required computing power equivalent to **about three hours of ChatGPT Pro "thinking"** .

Some of the claims are extraordinary. OpenAI says the model made progress on the **Kakeya conjecture**, advanced algorithms, and even touched on the **Riemann hypothesis** — one of the seven Millennium Prize Problems .

**Only 10 of the 722 manuscripts include reasoning summaries** explaining how the model reached its conclusions. And just **162 of the papers** — about **22%** — come with computer-checked formalizations in Lean, a programming language that mechanically verifies every logical step .

OpenAI itself admits: "some of the unformalized results could have issues" .

In other words, a lot of what was published might be wrong.

---

## The Mathematicians' Dilemma: Excitement vs. Skepticism

The reaction from the mathematical community has been a mix of astonishment, curiosity, and deep concern.

**The Skeptics: "Ask for Receipts"**

Andrew Sutherland, a mathematician at MIT, was blunt: **"Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified"** .

He added: **"We should ask for receipts"** .

Daniel Litt, a mathematician at the University of Toronto, was more excited about the potential but shared the concern about process. **"I think that it's great to have new solutions to questions that I and others are interested in,"** he told Fortune. But he worried about the effect on the field if AI is seen as having "solved math" .

**The "Proof Indigestion" Problem**

Terence Tao, one of the most respected mathematicians in the world, coined a phrase that captured the problem perfectly: **"proof indigestion"** .

The idea is simple. Machines can produce mathematical arguments faster than humans can verify, understand, or absorb them. And a proof that no one understands is, in a sense, not really knowledge — it's just a pile of symbols.

Tao wrote on social media: **"Problems are being solved autonomously by AI prompters who have no interest in the broader field itself once their initial target is 'solved', and do not understand the AI output well enough to answer questions on the result, give talks, or otherwise interact with the rest of the field"** .

**The "Slop" Concern**

The Association for Human Mathematics didn't mince words. In a statement criticizing the mass release, it declared: **"Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power"** .

The group urged mathematicians to stop collaborating with OpenAI and defend research centered on human understanding.

---

## The Core Problem: Translation and Trust

Here's where the technical issues get thorny.

When AI models solve math problems, they typically produce two things: a **natural language explanation** (the "proof" as written in English) and a **formal verification** in Lean (a machine-checkable code that proves the logic step-by-step).

The assumption is that if the Lean code compiles, the proof is correct. But a new paper by mathematicians at Cambridge and King's College London documented **discrepancies** between OpenAI's natural language proof of a Navier-Stokes-related result and its Lean code .

Specifically, they found that the Lean code proved a **weaker statement** than what the paper claimed. The paper called it "lost in translation" .

**This doesn't necessarily mean the proof is wrong.** But it raises a fundamental question: can we trust AI models to verify their own work? If the model can make mistakes translating its own ideas into code, what else might be slipping through?

Melanie Wood, a Harvard mathematics professor and member of the advisory group, put it simply: **"There is not human understanding of them at the point of release, and now the work begins"** .

---

## Frequently Asked Questions

**Q: What exactly did OpenAI release?**

A: OpenAI published **722 mathematical manuscripts** on GitHub, grouped into **372 result families**. The papers cover topics like algebra, theoretical computer science, geometry, and mathematical physics. They were produced by an internal, unreleased AI model that attempted roughly 4,000 problems .

**Q: Did OpenAI solve the Riemann hypothesis?**

A: **No.** OpenAI did not claim to solve the Riemann hypothesis. The release includes progress "toward" it and other major problems, but no complete solution to a Millennium Prize Problem has been verified .

**Q: Why are mathematicians skeptical?**

A: Three main reasons: (1) **Transparency** — OpenAI hasn't released the model or the prompts used, so results can't be replicated. (2) **Verification** — only about 22% of the papers have computer-checked formalizations in Lean. (3) **Understanding** — even verified proofs may be so opaque that no human understands them, which defeats the purpose of mathematical research .

**Q: What is "proof indigestion"?**

A: The term, coined by Terence Tao, describes the problem of machines producing mathematical arguments faster than humans can verify, understand, or absorb them. It's a problem of volume, not necessarily of quality .

**Q: What is Lean?**

A: Lean is a programming language used to formalize mathematical proofs. When a proof is translated into Lean, a computer can check every logical step mechanically. But passing a Lean check doesn't guarantee the proof is correct — it only shows the formalized version is valid. The formalization itself might contain errors .

**Q: What did the "lost in translation" paper find?**

A: Mathematicians at Cambridge and King's College London found discrepancies between OpenAI's natural language proof of a Navier-Stokes-related result and its Lean code. The Lean code proved a weaker statement than what was claimed in the paper. They say this shows that AI-generated proofs shouldn't be trusted without peer review .

**Q: Are any of these results actually breakthroughs?**

A: Some may be. Daniel Litt of the University of Toronto said he was excited about several results and eager to understand them. Abhishek Saha of Queen Mary University of London called it "a very big day for mathematics," though he noted most of the problems were "exceptional advances within an existing program" rather than game-changing breakthroughs .

**Q: What does the Institute for Advanced Study say?**

A: The Institute for Advanced Study in Princeton, which hosts the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), issued a statement saying it does not endorse OpenAI's practice of testing advanced problems on proprietary models. It said: **"It is now the case that AI can output mathematical arguments in situations without the human who prompted it being able to understand the arguments, verify them, or take responsibility for them"** .

**Q: What should I take away from this?**

A: AI is making genuine progress in mathematics, but the claims require careful verification. The technology is neither good nor bad — the outcomes depend on human choices about oversight and transparency. For now, treat unverified AI-generated proofs with caution .

---

## Conclusion: The Jewels and the Slop

Here's what I keep coming back to when I think about Andrew, the professor with 722 papers on his laptop.

He's not a Luddite. He's not anti-AI. He's a mathematician who cares about understanding — about *why* something is true, not just *that* it is.

OpenAI has produced something remarkable. Buried in those 722 manuscripts are likely genuine discoveries — solutions to problems that have stumped humans for years. Daniel Litt called them "great for mathematics." The Kakeya conjecture progress alone would be a career highlight for any researcher.

But there's also "slop" — unpolished, unverified, possibly wrong. OpenAI itself admits as much.

The real question isn't whether AI can solve math problems. It clearly can. The question is whether the mathematical community can absorb what AI produces without losing the thing that makes mathematics meaningful: **human understanding**.

Terence Tao's "proof indigestion" captures the problem perfectly. We've built a machine that can produce mathematical arguments faster than we can digest them. And a proof no one understands isn't knowledge. It's just symbols.

The jewels are there. The question is whether we can find them — and whether we'll still be able to understand them when we do.

---

## Disclaimer

**This article is for informational and educational purposes only. It does not constitute investment, financial, or academic advice. The author has no position in OpenAI or any related securities. Information presented here is based on publicly available sources and reported statements as of the publication date. The mathematical claims described have not been fully independently verified. Readers should consult official OpenAI communications, academic publications, and qualified mathematicians for the most current and accurate information.**