A Finnish logician says Anthropic's Claude helped him pull 663 theorems and proof ideas out of Kurt Gödel's private mathematical workbooks, a job his team had been doing by hand at roughly a page a week since 2017. Jan von Plato told The Guardian, in a story published November 13, that about 400 of those statements were verified as stated in the Coq proof assistant. The figures here are von Plato's, as relayed in coverage of the Guardian interview by explainx.ai and Superpower Daily.
Why it matters: this isn't an AI proving open problems. It's AI being used as a research assistant on a huge, hard-to-read archive, with a formal proof checker as the safety net. That combination — a language model to read and organize, a verifier to catch its mistakes — is the pattern many mathematicians expect to matter most. The episode also shows the limits: a proof checker can say whether a statement is true, not whether Gödel actually proved it.
The archive
Gödel kept workbooks in Gabelsberger, an obsolete German shorthand. The reports describe 16 workbooks, 1,009 pages, from the early 1940s. The University of Helsinki's Gödel Enigma project, which von Plato works on, lists the Arbeitshefte series as 16 notebooks of about 1,200 shorthand pages written mostly from the mid-1930s to the mid-1940s, so the exact page count depends on what's included. The project page lists von Plato's work as transcribing and translating these notebooks, and one of its stated aims is to develop Gödel's unpublished results on the axiom of choice and the continuum hypothesis.
What Claude did — and didn't do
According to the coverage, Claude did not read the shorthand. Von Plato gave it his own transcriptions, his earlier books and other scholarship on Gödel. From that it produced a 650-page digest in a couple of days, which von Plato says he checked page by page, and extracted the 663 candidates — everything from complete arguments to sketches and false starts.
The verification numbers
| Result after Coq checking | Count (von Plato's figures) |
|---|---|
| Held as stated | ~400 |
| Needed repairs (sharper statements or extra assumptions) | ~200 |
| False | ~50 |
A false result isn't necessarily an AI error. Superpower Daily notes Gödel sometimes marked his own pages as incorrect, so dead ends are part of the record.
The headline claim: the axiom of choice
The most eye-catching claim concerns the independence of the axiom of choice, a result later proved by Paul Cohen. Von Plato says Gödel came close but used one wrong definition, even though he had used the correct one two years earlier. He says Claude's reconstruction of the proof, using the correct definition, passed Coq.
Joel David Hamkins, a logician at Notre Dame, had not seen the reconstruction. He called it remarkable and historically interesting if it holds up, but said that if it resembles Cohen's approach it would show Gödel was on the right path — not that he had the proof — and that historians will need to separate Gödel's contribution from Claude's.
What isn't established yet
- No paper or data release has been reported, and explainx.ai notes no model version or cost was named.
- No outside review of the reconstruction or the 663-item list.
- Attribution — Coq verifies statements; it can't say whether Gödel held a given proof.
- Von Plato's broader claims about Gödel's standing are his own opinion.
What to watch: a publication of the digest or the Coq files would let other logicians check both the maths and the history. Until then, treat "400 verified" as the solid part, and "Gödel nearly proved independence" as a claim awaiting review.