ChatGPT’s 17-minute proof turns math into a verification problem

When Fields Medalist Timothy Gowers handed ChatGPT 5.5 Pro a set of open problems in number theory, he wasn’t testing whether the model could do math—he was testing whether it could invent it. The results, published in a preprint generated entirely by the AI, suggest it can. In under two hours, the model produced complete research papers, including a breakthrough that improved an exponential bound to a polynomial one in just 31 minutes and 40 seconds.

For context, that’s faster than most human mathematicians can even read a proof, let alone generate one.

The experiment, detailed in The Decoder’s report, wasn’t a cherry-picked demo. Gowers, who won his Fields Medal for work in functional analysis and combinatorics, gave the AI no mathematical guidance—just raw problems from a paper by number theorist Mel Nathanson.

The model’s output, evaluated by Gowers and MIT student Isaac Rajagopal, included a key idea Rajagopal described as 'completely original.' That’s not just a compliment; it’s a warning shot for human researchers. If an LLM can conjure original mathematical insights in minutes, what’s left for the rest of us to prove?

The source material also shows that the implications stretch beyond academia. Number theory isn’t just an abstract playground—it underpins cryptography, algorithms, and even blockchain security. If ChatGPT 5.5 Pro can autonomously improve bounds in these areas, it’s not hard to imagine AI-generated proofs being deployed in real-world systems, for better or worse. The MIT researcher’s praise for the model’s 'originality' is particularly striking.

Originality, after all, has been the last bastion of human intellectual superiority in math. If that’s now in play, the competitive landscape shifts dramatically.

Gowers’ takeaway is blunt: the new bar for mathematical contributions is 'proving something LLMs can’t.' That’s a high bar—and one that may only get higher as models like ChatGPT 5.5 Pro continue to evolve. The experiment also raises uncomfortable questions about verification. How do we trust a proof we didn’t write, especially when the AI’s reasoning isn’t fully transparent? The answer, for now, seems to be: we don’t. At least, not without a lot more scrutiny.

For the tech industry, this is a marketing milestone. For mathematicians, it’s an existential one. And for anyone who assumed AI would remain a tool rather than a colleague, it’s a wake-up call. The era of AI as a research partner isn’t coming—it’s already here.