OpenAI Has Your Unpublished Math. Should You Trust Them?
In this article
Bottom line: In late 2024, mathematicians learned that OpenAI had secretly funded Epoch AI's FrontierMath benchmark and received advance access to problems and reference solutions the broader research community hadn't seen — a fact that only came out after backlash forced disclosure.
That episode is back on Hacker News this week as researchers debate a newer wrinkle: contract language in AI math-research collaborations that lets labs retain unpublished proofs for "model improvement" indefinitely.
If you're a mathematician handing a lab your unsolved problem to test against GPT-5-class reasoning, you're trusting a company's internal data-handling policy over any enforceable norm the field has agreed on.
That trust has been broken once already, and the incentive structure that broke it hasn't changed.
I spent a decade keeping other people's secrets in production — API keys, customer data, the kind of stuff where a leak means an incident report and a very bad Monday.
So when I read mathematicians on Hacker News this week arguing about whether OpenAI can be trusted with an unpublished proof, I recognized the shape of the problem immediately. This isn't a math story.
It's an access-control story that happens to involve math.
And access-control stories have a predictable failure mode: the people granting access assume good faith, the people receiving it optimize for their own incentives, and nobody finds out where the line was crossed until it already has been.
The Setup: Why Mathematicians Are Even Talking to AI Labs
Here's the scenario that keeps recurring.
A frontier lab — OpenAI, Anthropic, Google DeepMind — wants to claim its model can do "real" mathematical reasoning, not just regurgitate Olympiad problems it's seen a thousand times in training data.
To prove that, they need problems the model has never encountered. Ideally, problems so new that no solution exists anywhere on the public internet.
That's an extremely narrow resource. The people who have it are working mathematicians sitting on unpublished results, half-finished proofs, or problems they're still actively researching.
So labs go recruit them — sometimes directly, sometimes through intermediaries like Epoch AI, which built the FrontierMath benchmark specifically to solve this problem for OpenAI in 2024.
The pitch to mathematicians is reasonable on its face: contribute a hard, original problem, get paid, help the field understand what these models can actually do. Dozens of contributors signed on.
What most of them didn't know — because it wasn't disclosed — was that OpenAI was funding the benchmark and had access to the problems and solutions before anyone else did. Epoch AI confirmed this in January 2025, after independent researchers noticed OpenAI's name buried in a footnote and started asking questions publicly.
The defense at the time was that a "held-out" subset of problems was kept from OpenAI.
Mathematicians who'd contributed problems pointed out, correctly, that they had no way to verify that claim, and no contractual mechanism forcing anyone to honor it.
The Core Insight: Trust Doesn't Scale the Way Data Does
Here's the part that infrastructure people understand instinctively and that keeps tripping up domain experts in other fields: once data leaves your control, your trust model is only as good as the weakest enforcement mechanism behind it.
In distributed systems, we don't say "trust the server to behave correctly." We say "assume the server can be compromised, malicious, or simply wrong, and design so that doesn't matter." Zero trust architecture exists precisely because "we promise not to misuse this" is not a control — it's a hope.
Math research has been operating on the hope model.
A mathematician sends an unpublished proof to a lab for evaluation, and the entire safeguard is a paragraph in a data-use agreement, enforced by nothing more than reputational risk.
That worked fine when the counterparty was another mathematician bound by the same career incentives as you.
It works very differently when the counterparty is a company racing competitors to the next benchmark headline and burning nine figures a month to do it.
The Contamination Problem Nobody Can Fully Audit
The specific fear driving the current Hacker News thread isn't just "OpenAI might read my proof." It's that an unpublished problem, once ingested by a lab in any form, might end up — deliberately or through ordinary data pipeline sloppiness — in a future training run.
If that happens, the next model doesn't "solve" your problem.
It memorizes the answer it was shown, and then a benchmark score gets reported as evidence of reasoning capability that was actually just leakage.
This isn't hypothetical paranoia. Benchmark contamination is a well-documented, recurring problem across every major lab, going back to GPT-4-era complaints about coding benchmarks.
The difference with unpublished math is that there's no public source to compare against.
If a model suddenly gets unusually good at a class of problems whose only prior appearance was a private submission to that same lab, there is no external way to prove contamination occurred. You'd need the lab's training logs, which nobody outside the lab has ever seen.
Why "We Have a Policy" Isn't an Answer
When this came up in 2025, OpenAI's response leaned on internal data-handling commitments — segregated access, contractual restrictions on training use.
Reasonable people can debate whether those commitments were honored.
But the structural problem is that the policy is unauditable by the people it's supposed to protect. A mathematician has no tooling equivalent to a SOC 2 report, no way to request a data-access log, no third party attesting that the held-out set stayed held out.
Compare that to how a security-conscious engineering org handles a similar problem — vendor access to sensitive source code. You don't take their word for it.
You demand SSO-gated access logs, time-boxed credentials, and a contractual audit right.
Math research collaborations with AI labs have none of that tooling, because the field never needed it before language models made "give us your hardest unpublished problem" a business model.
The Reality Check: This Isn't Purely a Villain Story
It would be easy to write this as "OpenAI bad, mathematicians victims," and plenty of the angrier Hacker News comments do exactly that.
I don't think that's the full picture, and I'd rather be right than satisfying.
Building genuinely uncontaminated math benchmarks is a real and difficult problem, and someone has to fund it.
Epoch AI needed money to pay mathematicians for their time, and OpenAI wasn't the only lab that benefited from FrontierMath's existence — Anthropic and Google have both cited it in model cards.
The failure wasn't collaboration. It was the lack of disclosure, and the absence of any enforcement mechanism once trust was extended.
It's also worth saying plainly: no evidence has surfaced that OpenAI actually trained on the held-out FrontierMath set.
The scandal was about undisclosed access and conflict of interest, not proven contamination. That distinction matters, and it gets flattened in outrage threads.
The trust problem is real even without a confirmed leak, because the arrangement made a leak undetectable either way — and undetectable is its own kind of failure.
The Practical Takeaway: Treat It Like a Vendor Security Review
If you're a researcher — math, physics, anything with a similar dynamic around unpublished results and AI evaluation — stop treating these collaborations like academic favors and start treating them like vendor contracts, because that's what they are.
Concretely:
- Demand disclosure of every funding relationship behind a benchmark or evaluation program before you contribute anything unpublished, not after.
- Push for time-boxed, revocable access to your material, with a contractual right to request deletion confirmation — the same terms you'd expect from any SaaS vendor touching sensitive data.
- Publish your problem's hash or a redacted description publicly before submission, so you have a timestamped, independent record if a suspiciously similar solved problem shows up in a future model's marketing materials.
- Assume anything you submit privately may eventually become public capability, and only submit what you're genuinely fine losing control of.
None of this requires trusting anyone's intentions. That's the point — it doesn't have to.
Security engineers gave up on trusting good intentions decades ago, which is why zero trust architecture exists at all.
Math research is now running the same experiment the software industry already ran, just a decade behind and with proofs instead of production databases.
So: if a lab reached out tomorrow asking for your hardest unsolved problem, what would you actually need to see in writing before you said yes — and do you think that document exists anywhere right now?
