OpenAI is making headlines again, this time with claims of three major breakthroughs: a proof of the local smoothing conjecture in mathematics, a refutation of the Exponential Time Hypothesis in computer science, and a mathematically rigorous existence proof for certain quantum field theories in physics. But what strikes me most isn’t the breakthroughs themselves, it’s how the company is handling their release and the uncomfortable questions this raises about the future of research validation.
Just last month, OpenAI released hundreds of AI-generated mathematical manuscripts. This time, they’re claiming to have learned something from feedback, bringing in external experts as “shepherds” to guide verification. For the computer science result, they’ve enlisted Russell Impagliazzo and Ramamohan Paturi, who actually proposed the Exponential Time Hypothesis back in 1999. That’s clever optics, but it also highlights something troubling: we’re now in a position where original conjecture authors need to verify claims that contradict their life’s work.
The Verification Problem Nobody’s Talking About
Here’s what bothers me about the current approach. OpenAI is essentially asking experts to become gatekeepers of AI-generated results. The company released an open letter asking the community how to handle “early access to AI-generated results” and “conflicts of interest.” That’s a loaded question to pose after already having designated specific people as shepherds.
The AGMAI panel of math advisers that OpenAI assembled literally just criticized this exact practice yesterday. They warned against giving certain people first access to results, citing unequal access, borrowed reputation, and biased selection. Yet OpenAI is doing exactly that with the quantum field theory work, where David Gross and colleagues get to “shepherd” the physics results.
There’s a fundamental asymmetry here that developers and researchers need to understand. When human mathematicians publish proofs, peer review is about validating logic and catching errors. But when AI generates mathematical content, we’re asking reviewers to do something harder: they need to understand not just whether the proof is correct, but whether the AI actually discovered something novel or simply pattern-matched from training data in a new way.
What This Means for Research and Development
I’ve been following AI breakthroughs closely, and this moment feels significant. We’re at an inflection point where AI systems can produce outputs that look like peer-reviewed research but lack the epistemic transparency we typically expect.
For developers building on top of these results, this creates real uncertainty. The computer science claim about ETH gets released with Lean proof formalization and GitHub artifacts, which is good. That’s verifiable. But how many downstream projects will cite these results before independent verification is complete? What happens if the proofs contain subtle errors that take months to surface?
The open letter asks important questions about how credit should work and how we support researchers as “methods and research directions change.” I appreciate the framing, but I think it’s underselling the actual problem. This isn’t just about proper attribution or supporting students. It’s about whether we can maintain meaningful verification standards when AI can generate sophisticated-looking output faster than humans can validate it.
The Precedent We’re Setting
Consider the math result about local smoothing. OpenAI’s AI produced a proof that apparently resolves the Kakeya conjecture in all dimensions. Hong Wang just won a Fields Medal for resolving it in three dimensions. That’s a stunning difference in scope, and it raises obvious questions: Did the AI really solve a harder problem, or did it generate something that looks like a proof but isn’t one? We won’t know quickly.
What worries me most is the precedent. If companies can release AI-generated research and ask external experts to figure out what’s true, we’ve shifted from “verify before publication” to “publish and figure it out later.” That works fine for blog posts. It’s dicier for fundamental research that other scientists might build on.
I’m not arguing against AI-assisted research. I’m arguing that the current release strategy puts verification experts in an impossible position and creates perverse incentives. When OpenAI releases results for shepherds to “explain,” they’re essentially outsourcing validation to whoever they choose, which is exactly what the AGMAI panel warned against.
The real question isn’t whether AI can produce mathematical insights. It’s whether we’re ready to rebuild scientific verification practices for an era when machines can output research faster than humans can read it, much less validate it.