Claude Found 663 Gödel Theorems; About 400 Held Up in Coq

The AI model Claude extracted 663 theorems from Kurt Gödel’s 1940s workbooks, and when they were put into the Coq proof assistant about 400 were verified as stated, about 200 were repaired and about 50 were false, The Guardian reported on 13 November. The figures come from Jan von Plato, the Finnish logician who leads a project at Helsinki University on the 1,009 pages of notes.

That is about 60% verified as stated, about 30% repaired and about 8% false, by our arithmetic. At the project’s earlier pace of about a page a week, 1,009 pages is about 19 years of work; the Guardian reports that Claude produced a digest of the entire material within a couple of days, which von Plato said had to be checked page by page.

Business Pill · Verification Cost

An evergreen Business Engineer explainer on why output that is cheap to make is not cheap to check. It teaches the general idea only and does not discuss Gödel, Claude or this project.

The key insight: As we read the figures, the number that matters is the split, not the count: about 90% of the 663 statements Claude extracted held up in Coq as stated or after repair, and about one in 13 was false, by our arithmetic.

What the Guardian Reported

The Guardian’s Alex Bellos reported that von Plato has led the Helsinki project since 2017 to understand 16 notebooks Gödel wrote in the early 1940s. The material, the “workbooks”, is 1,009 pages on maths and logic, much of it in Gabelsberger, a German shorthand the Guardian says “was obsolete even when Gödel was using it”.

Von Plato received a €2.5m grant from the European Research Council, taught himself the shorthand and assembled a small team, the Guardian reported. Von Plato told the paper: “I am one of only six people who can read it in the world”.

In August, at a conference in France on AI and theorem proving, computer scientists suggested von Plato use Claude, the report says. Back in Helsinki, von Plato showed Claude “his transcriptions of the workbooks, as well as six books he had previously written on Gödel and other secondary scientific literature”. So Claude worked from von Plato’s transcriptions, as the Guardian describes it, not from the shorthand pages directly.

Von Plato described the result: “In a couple of days, I got to see a 650-page explanation of the workbooks. It had to be controlled page for page, but it’s basically a work that is impossible to do for human beings”.

Outcome in the Coq proof assistant for the 663 theorems Claude extracted from Gödel’s workbooks, as appr
Outcome in the Coq proof assistant for the 663 theorems Claude extracted from Gödel’s workbooks, as approximate counts given by Jan von Plato to The Guardian: about 400 verified as stated, about 200 repaired, about 50 false. Shares of 663 are our arithmetic; the rounded counts add to about 650.

What Coq Checked

Von Plato asked Claude to extract every theorem in the workbooks for which Gödel “either had a proof, had sketched a proof or had an idea for a proof”. Claude found 663, the Guardian reported. Von Plato described the Coq results: “About 400 were verified as stated, about 200 were repaired, meaning that perhaps the statement of the theorem had to be made more precise, or some assumption had to be added. About 50 were false, and this is no wonder”.

Von Plato added that Gödel, rereading the notes, would occasionally write “falsch!” or “schlecht!” in the margin. About 400 plus about 200 is about 600 of 663 statements that held up as stated or after repair, about 90%, and about one in 13 was false, by our arithmetic. The three rounded counts add to about 650, so roughly 13 of the 663 are not covered by them.

The Guardian does not say who put the 663 theorems into Coq. For one proof it does: von Plato ran Claude’s reconstruction of Gödel’s proof on the axiom of choice through Coq, “a program that verifies the validity of mathematical proofs”, and it passed, the report says.

The Claim Gödel Made About Cohen

In 1966 the American mathematician Paul Cohen won the Fields medal for showing the independence of the continuum hypothesis and the axiom of choice, the Guardian notes. Gödel reacted by saying the result had already been proved in the 1940s, but could not reconstruct a proof from the notes when peers asked for one, the report says.

The first task von Plato set Claude was to look for that proof in the workbooks. The answer, per the Guardian, was “yes, almost”. Von Plato said: “He was just one step away”, adding that “There’s one definition that Gödel got wrong, but he had the right definition two years earlier”.

Joel David Hamkins, a professor of logic at the University of Notre Dame, has not seen Claude’s reconstruction, the Guardian reported. Hamkins said that if it stood up “it is remarkable and historically interesting. Although how much is Claude and how much is Gödel is a job for the historians”. Hamkins also said: “You don’t get credit for work you don’t tell anyone about”.

From a Page a Week to a Couple of Days

Progress in the mathematical understanding of the text was “painfully slow – about a page a week – until September”, the Guardian reported, when von Plato enlisted Claude. Von Plato put the human limit this way: “It takes us a year to read about 50 pages”.

At about a page a week, the 1,009 pages would take about 19 years; at about 50 pages a year, about 20 years, by our arithmetic. Taking “a couple of days” as two, Claude’s digest ran at about 500 pages a day against about one-seventh of a page a day, roughly 3,500 times the earlier pace, by our arithmetic. That digest was 650 pages and still had to be checked page by page, von Plato said.

Von Plato told the paper: “The AI can connect all the pieces in a thousand pages. It can see all of Gödel’s writings as one piece.” Von Plato also said the workbooks contain, “Claude says, more than a dozen themes in logic and foundations of mathematics that have not been found in the 80-plus years since they were written down”.

Time to read Gödel’s 1,009-page workbooks: about 19 years at a page a week, about 20 years at 50 pages a year, versus Claude’s digest in a couple of days
Reading paces as described by Jan von Plato to The Guardian; the 19-year and 20-year figures are our arithmetic (1,009 pages at about a page a week and at about 50 pages a year). The Claude bar is not to scale.

Where This Sits in Our AI-for-Science Coverage

The Guardian’s report adds a historical archive to a run of AI-in-research stories we have covered: OpenAI’s new science blog, Terence Tao’s six proposals for maths after OpenAI’s proofs and DeepMind’s Pushmeet Kohli on a phase transition in AI science. In this case the checking step is public in the numbers: about 400 of 663 statements passed Coq as stated, per von Plato.

The Structural Read

As we read the Guardian’s account, the work moved from reading to checking: Claude produced a 650-page explanation from von Plato’s transcriptions within a couple of days, which von Plato said had to be controlled page for page; the extracted theorems were later put into Coq, the report says.

As we read the Cohen episode, the evidence is a reconstruction, not a recovered proof: von Plato said Gödel got one definition wrong, Claude’s reconstructed proof passed Coq, and Hamkins, who had not seen it, said how much is Claude and how much is Gödel is for historians.

Jan von Plato, to The Guardian

“The AI can connect all the pieces in a thousand pages. It can see all of Gödel’s writings as one piece.”

Three Implications

FOR RESEARCH TEAMS The Guardian reports the shift came from a conference suggestion in August, nine years into the project, by our arithmetic; the 650-page digest still had to be checked page by page, von Plato said.

FOR VERIFICATION About 400 of 663 statements were verified as stated, about 200 repaired and about 50 false in Coq, per von Plato: about 60%, 30% and 8%, by our arithmetic.

RELATED Our piece on Terence Tao’s proposals for maths after OpenAI’s proofs covers how mathematicians are setting rules for AI-assisted results.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Product Overhang Doctrine.

The framework argues: “The first is what your underlying technology can do. The second is what your users can actually do with it.”

As we read it, the Gödel project shows that gap in a research setting: per the Guardian, von Plato began using Claude in September after computer scientists suggested it at an August conference, and the material then moved from about a page a week to a digest within a couple of days. The framework predates this story and does not discuss it.

What Is Not Established

The Guardian’s article names Claude but does not mention Anthropic, the company that makes it, or say which Claude model was used. It does not cite a paper or preprint, and we did not find one on arXiv when we searched on 16 November. All the counts are approximate figures given by von Plato; we did not see the Coq files.

Hamkins, the only outside logician quoted, had not seen the axiom-of-choice reconstruction. Whether Gödel had Cohen’s result is a question von Plato and Hamkins frame differently in the report. We did not contact von Plato, Helsinki University or Anthropic for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Product Overhang Doctrine

A Business Engineer framework on the gap between what AI models can do and what users actually do with them.

Read the premium analysis →Get 50% off →

The Bottom Line

The Guardian reported that Claude extracted 663 theorems from Gödel’s 1,009 pages of workbooks, working from von Plato’s transcriptions, and that in Coq about 400 were verified as stated, about 200 repaired and about 50 false: about 60%, 30% and 8%, by our arithmetic. At the project’s earlier page-a-week pace, the same pages would have taken about 19 years.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every fact about von Plato’s project, the 663 theorems, the Coq counts, the axiom-of-choice reconstruction and the quotes from von Plato and Joel David Hamkins is from The Guardian’s 13 November article by Alex Bellos, which we read in full. The Coq counts are approximate figures given by von Plato; we did not see the underlying files, and the Guardian cites no paper.

The story reached us through a post on X that relays the Guardian. The shares of 663, the 19-year and 20-year reading times and the pace comparison are our arithmetic. Von Plato, Helsinki University and Anthropic were not contacted for this piece.

Sources: The Guardian, 13 Nov 2026 (Alex Bellos)

Scroll to Top
FourWeekMBA