Fields Medalist Tests ChatGPT 5.5 Pro: PhD-Level Math Research in Under Two Hours
Fields Medalist Timothy Gowers published a widely-discussed blog post documenting his experience with ChatGPT 5.5 Pro on mathematical research. The result is arresting: with essentially no mathematical input from Gowers, the model produced PhD-level original mathematics in just over an hour.
How it worked
Gowers selected a set of open problems from a Mel Nathanson paper on additive number theory. The core question: given a set of integers of size k, what sizes can its h-fold sumset take?
After 17 minutes and 5 seconds of reasoning, ChatGPT 5.5 Pro produced a construction that improved an exponential upper bound to a quadratic one — the best possible result. Gowers then asked it to write up the proof as a LaTeX preprint, which took another 2 minutes and 23 seconds.
More impressively, when Gowers asked the model to improve on prior work by MIT undergraduate Isaac Rajagopal, ChatGPT came up with what Rajagopal himself called “an original and clever idea” — the kind he’d be proud to come up with after a week or two of pondering. The result was verified as correct.
What it means
Gowers raises several important questions:
The bar for PhD training has shifted. If LLMs can solve “gentle” open problems, new researchers can no longer start by tackling easy open questions. The new baseline: prove something LLMs can’t prove, or collaborate with LLMs to go beyond what they can do alone.
Attribution becomes fuzzy. The ChatGPT-generated proof would be publishable if produced by a human. But who gets credit? Gowers contributed zero mathematical input. He suggests a separate repository for AI-generated results with human verification.
The pace is accelerating. Gowers notes that everything he describes is based on current capability — and that these comments will likely be “out of date in a matter of months.”
Implications for AI and knowledge work
This isn’t just a math story. It demonstrates that frontier LLMs are crossing the boundary from “assistive tool” to “independent researcher.” For industries relying on AI for knowledge work, this means workflows, quality assessment, and trust mechanisms need redesign. When AI can not only execute tasks but discover new knowledge, the human role in the knowledge production chain changes fundamentally.