AI Disproves 27-Year-Old Hypothesis, Leaving Computer Scientists Shaken

What this is about

This week, OpenAI and Anthropic both announced that they had disproved the Exponential Time Hypothesis (ETH). The ETH is a 27-year-old assumption in theoretical computer science. It says that the effort needed to solve certain computational problems grows exponentially as the problems get bigger. Exponential means that every small step up in problem size multiplies the effort. WIRED and New Scientist reported on it, and other outlets followed.

A lot of research on algorithms and complexity (the question of how hard computational problems fundamentally are) builds on this assumption. A contradiction could therefore invalidate entire lines of research. Important: the proofs have not yet been independently checked. For now, these are the companies' claims.

What researchers fear

Peter Jonsson, a professor at Linköping University, is due to present a paper at the IEEE Symposium on Foundations of Computer Science in two weeks. A co-author will deliver the talk. According to WIRED, much of the paper has been "dead" since Monday. He expects many others to be in the same boat: "People will start their talks with: 'ETH is false, half of my paper doesn't work.'" He thinks it may be possible to rescue his work with a different method, though.

Daniel Lokshtanov (UC Santa Barbara) says that if the result is confirmed, he was "galactically wrong" and about 50 of his publications would become "meaningless" or need revising. He calls the result a "huge bomb in the middle of the field" – "really cool and terrible at the same time."

Thore Husfeldt, head of the algorithms group at the IT University of Copenhagen, had spent almost the entire year thinking about a single problem. "As far as I can see, it's all irrelevant now." He had formulated a version of the hypothesis that is now supposedly disproved, and he is working out on a whiteboard which of his papers are affected. Still, he calls the result "fantastic" and a "result of the century."

Checking the proofs

Both companies have published their proofs on GitHub: one in natural language, one in Lean. Lean is software that lets computers check mathematical proofs step by step. But a formally correct proof isn't the same as one that the research community can understand and build on.

OpenAI has asked the researchers who originally proposed the hypothesis, along with an MIT scientist who disproved a version of the ETH, to organize the review. WIRED initially misidentified the originators as being from UC San Diego and later corrected this. Other reports say UC Santa Barbara, so attribution is inconsistent across coverage. So far there is a Google Form for volunteers and OpenAI's pledge to fund a conference. Lokshtanov plans to take part. According to WIRED, OpenAI's roughly 60-page paper is better written than the companies' earlier publications.

A flood of AI results and the toll on mental health

Researchers complain about the sheer volume of AI-generated results. Earlier this month, OpenAI uploaded 722 manuscripts covering 400 problems to GitHub, to the annoyance of academics who can hardly review that much material. Some documents contained errors. According to WIRED, some mathematicians saw it as a rushed move meant to help the company's planned IPO.

In the middle of an interview, Husfeldt received a 17-page paper, probably AI-generated, on the consequences of the ETH from an unknown Gmail address. He didn't read it: "Reading math is a very big investment of cognitive time." He says: "The last few months have been really hard," because AI results "destroy" ongoing projects almost every week. His department has considered hiring an occupational psychologist, since some colleagues are "really depressed."

Implications for cryptography

Encryption relies on the assumption that certain problems can't be solved in practical time. According to Lokshtanov, the ETH result doesn't directly affect existing cryptography, "but it should definitely scare some cryptosystem designers." Husfeldt stresses that the open scientific process should expose any misunderstandings about computation, since a lot of digital infrastructure rests on such assumptions.

Opportunities

Victor Lagerkvist (Linköping) wants to return to problems he abandoned on the assumption that the ETH held. Jonsson says a verdict on the ETH is something he could only "dream" of: "Whether a human wrote the proof, an AI or aliens – if it's right, we have to use it."

Background: AI in mathematics

A September 2026 piece from Emory University shows the context. OpenAI had previously said it used AI to solve the regularity problem for the Navier-Stokes equations, a Millennium Problem with a million-dollar prize. That sparked debate about roles, attribution and professional ethics. Mathematician Levon Nurbekyan uses AI himself for literature searches, prototyping and as a source of ideas, but warns about hallucinated sources and the loss of social exchange in research.

Our take

If it holds up, this would be one of AI's biggest contributions to theoretical computer science. For now, though, it is an unconfirmed claim from two companies with their own interests. It remains to be seen whether the proofs survive scrutiny and how far the consequences reach for textbooks and cryptographic assumptions. Notably, even those affected see the result more as an opportunity, while the flood of AI papers and the mental strain are real problems for the field.