OpenAI's "solved" problems count for nothing until experts say so

When a company says its AI solved problems that stumped researchers for decades, what should we believe? Our answer: nothing yet. Not because it's impossible, but because in science an announcement is not a result. A result is something others can check, break and reproduce.

What OpenAI is claiming

According to the coverage we reviewed, OpenAI announced advances on three open problems in mathematics, theoretical computer science and physics. One of its internal models, still unreleased, allegedly refuted the Exponential Time Hypothesis (ETH), a conjecture that the community had long accepted as true, as Medellín Style summarizes. In a related story, Wired reports that Anthropic and OpenAI say they have disproved a key tenet of computer science, and that researchers are struggling with the fall but also feel like "overexcited bunnies."

The part that matters most is the process. 20minutos reports that outside experts, among them Nobel laureate David Gross, will examine some of the work. As Gizmodo's headline frames it, OpenAI now needs to convince the experts that it's true.

Verification is the product, not a footnote

The excitement is understandable. If a model really did refute ETH, that would be a big deal. But "refuted" is a heavy word, and it only becomes true once specialists have gone through the argument line by line. Until then, we have a company with every incentive to be impressive telling us it was impressive.

One of the secondary sources we read makes a point we find sharper than the headline. Ecosistema Startup argues that OpenAI changed how it publishes because the old approach cost it credibility. In its account, the company moved from a mass dump of results to a scheme in which external experts coordinate the interpretation, and in physics those experts decide not to publish the AI-generated materials. That is one outlet's reading, not a confirmed account of the company's motives. But it matches what we'd expect: after losing credibility, the cheapest way to earn it back is to hand the microphone to someone else.

We think that's the right instinct, executed halfway. Having reviewers is good. Not publishing the physics material is not, at least not without a clear explanation of why it can't be shared. Science doesn't run on "trust our panel."

What OpenAI owes us

Three things would turn this from a press release into a contribution. First, the proofs, in full, in a form mathematicians can read and, where possible, check formally. Second, the method: what the model was asked, how many attempts it took, how much human steering went in, and how failures were filtered out. Third, the unreleased model's role: if nobody outside can run it, reproducibility is partly theoretical.

The method matters as much as the answer. A proof that is correct but was extracted by dozens of human nudges is a different story from one the model produced alone. At least one aggregator, gptzone.net, already describes the results as generated almost entirely by the AI, but that is an enthusiastic outlet's wording, not verified documentation, and we wouldn't lean on it.

Our verdict

Watch, don't applaud. If the experts confirm the work, OpenAI deserves real credit, and the field will need to rethink what counts as hard. If they don't, this will join the long list of AI triumphs that shrank on contact with peer review. Either way, the lesson for readers is the same: the headline "AI solves decades-old problem" should be read as "AI company asserts." The verification, whenever it arrives, is the actual news.

For OpenAI, the path is simple and unglamorous: publish everything, invite attacks, and let the critics be loud. A discovery that can't survive that was never a discovery.