Tobias Reithmeier

Blog

Claude Reads Gödel's Workbooks: 663 Theorems and an Almost Finished Proof

The Finnish logician Jan von Plato has put the AI model Claude to work on 16 workbooks written by Kurt Gödel. Claude found 663 mathematical theorems in them for which Gödel had noted a proof, a proof sketch or an idea for a proof. About 400 passed the proof assistant Coq as recorded, about 200 had to be repaired and about 50 were false. The Guardian reported it first; I came across it through the German site all-ai.de.

The story sounds like an AI discovering lost mathematics. Read closely, it is more sober: a model reads a thousand pages in one go, a human checks every page, and a program verifies the proofs. So far the only evidence is what von Plato told a newspaper. I find the history behind it more interesting than the numbers. It starts with a school shorthand from the 19th century and leads to a claim by Gödel that nobody has been able to check for almost sixty years.

Who wrote the workbooks

Kurt Gödel was born in 1906 in Brünn, today's Brno, and studied in Vienna. In 1931 he published the two incompleteness theorems. Simplified, they say that every consistent formal system strong enough for arithmetic contains statements that can be neither proved nor refuted within it, and that such a system cannot prove its own consistency. That ended the hope of a generation that all of mathematics could be derived from one fixed set of rules. Alan Turing built on it in 1936 when he laid the foundations of computability.

In January 1940 Gödel left Vienna with his wife Adele. Crossing the Atlantic was too risky in wartime, so they took the Trans-Siberian Railway east, went on to Japan and from there by ship to San Francisco. They reached Princeton in March. He stayed at the Institute for Advanced Study, with which he had been associated since 1933, until his death in 1978. There he became a close friend of Albert Einstein, and the two walked home together every day.

After the great papers of the 1930s Gödel published very little. He kept writing, though: tens of thousands of pages that he never prepared for print. The 16 workbooks date from his first years in Princeton and hold 1,009 densely written pages on logic and the foundations of mathematics. The papers belong to the Institute for Advanced Study and have been on deposit at Princeton University's Firestone Library since 1985.

A script almost nobody reads any more

Much of the workbooks is written in Gabelsberger shorthand. Franz Xaver Gabelsberger, a clerk in Munich, developed it from 1817 and published it as a textbook in 1834. It became the most widely used stenography in the German-speaking countries and was taught in schools. According to the Helsinki research group, Gödel learned it at school between 1919 and 1921; his textbooks are among his papers. In 1924 the unified German shorthand, the Deutsche Einheitskurzschrift, replaced the system. When Gödel filled his workbooks in Princeton he was already writing in a script on its way out, and he kept to it for the rest of his life.

Shorthand drops vowels, reduces frequent words to single signs and depends on the writer's hand. The Helsinki group describes reading it as a mix of rules, understanding of the content and a good deal of guesswork. If you do not know which theorem of logic could be meant at a given point, you often cannot resolve the sign for it. That is why these workbooks can only be transcribed by people who know both stenography and mathematical logic.

The project in Helsinki

Jan von Plato is one of them. According to the Guardian he started on the notebooks in 2017 because he was facing mandatory retirement at the University of Helsinki and wanted to make himself indispensable. He taught himself the shorthand and counts himself among six people in the world who can read it. The European Research Council funded the work under the name GODELIANA. The Guardian gives the grant as 2.5 million euros; the Finnish research portal lists about 2.13 million euros for 2018 to 2024.

The project has already produced books, among them an edition of Gödel's shorthand notes on incompleteness and chapters from a book on foundational research that Gödel never finished. The workbooks remained the largest unexplored source. The transcription into ordinary script existed, but the mathematical understanding advanced slowly, by von Plato's account at about a page a week.

How Claude came in

In August 2026, at a conference in France on AI and theorem proving, von Plato met computer scientists who suggested bringing in Claude. Back in Helsinki he gave the model his transcriptions of the workbooks, six of his own books on Gödel and further scholarly literature.

A couple of days later he had a 650-page explanation of the workbooks. Von Plato says two things about it that belong together. First, he had to check the analysis page by page. Second, he considers it work that humans cannot do in this form: someone who reads one page after another and needs a year for 50 pages no longer has the other 950 in mind. The model, he says, can connect all the pieces across the thousand pages.

It matters what Claude did not do. The team transcribed the shorthand over years, and Claude worked with those transcriptions. Headlines claiming that an AI deciphered Gödel's shorthand misstate the report.

663 theorems and what Coq left of them

Von Plato asked Claude to collect every theorem in the workbooks for which Gödel had a proof, a sketch or an idea. The result went through Coq.

Coq is a proof assistant that was developed from 1984 at the French research institute Inria and has been called Rocq since 2025. A program of this kind is not a tool for computing results and usually does not search for proofs on its own either. It checks step by step whether a formally written proof follows from its assumptions, and it lets no gap pass that a human reader would skim over. Coq became widely known in 2005, when Georges Gonthier used it to verify the four colour theorem completely, a theorem whose original computer proof of 1976 had long been disputed.

Result in CoqCountMeaning
verified as statedabout 400The theorem held in the form in which Claude had recorded it
repairedabout 200The statement had to be made more precise or an assumption added
falseabout 50The theorem does not hold

The false theorems do not surprise von Plato. Rereading his notes, Gödel himself sometimes wrote "falsch!" or "schlecht!" in the margin, wrong or bad. Workbooks are a workshop, not a textbook.

To judge the numbers it helps to say exactly what Coq checks. The program confirms that a formalised proof has no gaps. It cannot know whether the formalisation renders what Gödel meant in 1941. Between the note in the workbook and the verified proof lie two translations: the transcription from shorthand and Claude's version for Coq. At both points something can enter that is not Gödel's.

The axiom of choice: sixty years of dispute

The most striking single question in the workbooks concerns the axiom of choice, and understanding it takes some background.

The axiom says: for every collection of non-empty sets there is a rule that picks exactly one element from each of them. For finitely many sets that is obvious. For infinitely many it asserts the existence of a choice that nobody can write down. Ernst Zermelo formulated it in 1904 to prove that every set can be well-ordered, and started a dispute that lasted for decades. A large part of mathematics needs the axiom. At the same time it implies things like the Banach-Tarski paradox of 1924, according to which a ball can be cut into finitely many pieces and reassembled into two balls of the same size.

So the question was how the axiom relates to the other rules of set theory. Two answers were missing. Does it lead to a contradiction? And does it perhaps follow from the other axioms anyway?

Gödel himself gave the first answer. In 1938 he announced that the axiom of choice and the generalised continuum hypothesis are consistent with the other axioms. To show it he built a small universe of sets that can be defined step by step, the constructible sets, and proved that both statements hold in it. The full account appeared in Princeton in 1940, the year he arrived. That settled one direction: if the other axioms are consistent, they stay consistent with the axiom of choice added. So it cannot be refuted from them.

The other direction stayed open, the independence: can it not be proved either? That requires a universe in which all the other axioms hold and the axiom of choice is false. This is what Gödel was working on in the early 1940s, the time of the workbooks. He published nothing about it.

Paul Cohen found the solution in 1963. He was a young analyst at Stanford and not a logician. He invented a new method for it, forcing, which extends universes of sets in a controlled way, and proved the independence of the axiom of choice and of the continuum hypothesis. Gödel checked Cohen's work and submitted it himself to the US National Academy of Sciences for publication. In 1966 Cohen received the Fields Medal.

What Gödel claimed

After Cohen's award Gödel said he had had the result back in the 1940s. The Guardian puts it briefly. The record is more precise and more cautious. In letters to Alonzo Church in September 1966 and to Wolfgang Rautenberg in June 1967, Gödel wrote that he had found an independence proof for the axiom of choice in 1942 and gave his reasons for not publishing it. He meant type theory, a weaker framework than the set theory in which Cohen proved his result. Gödel's biographer John W. Dawson conjectures that the proof was found in the summer of 1942 in Blue Hill on the coast of Maine. Gödel's confidant Hao Wang pointed to the notebooks from that period.

Later Gödel could not reconstruct the proof from his notes. The matter has counted as unresolved ever since: nobody knows whether he had a complete proof in 1942 and merely held it back. When his correspondence appeared in the Collected Works in 2003, the question of what the notes contain was still open.

This is where von Plato's first question to Claude comes in: do the workbooks contain such a proof? The answer was: almost. Gödel's draft contains one faulty definition. He had written down the right one two years earlier. Von Plato thinks Gödel would have arrived at it himself with a few more months of work. He ran the version Claude reconstructed from this through Coq, and it passed.

Von Plato also believes that Gödel would have won the 1940 Fields Medal had he published. There was none: after 1936 the medal was not awarded again until 1950, because of the war. That changes little about the core of his assessment, but it shows that this is someone telling the story with enthusiasm, not with caution.

What remains open

The clearest reservation is in the Guardian article itself. Joel David Hamkins, professor of logic at the University of Notre Dame, has not seen the reconstruction. If it stands up, he calls it remarkable. How much of it is Claude and how much is Gödel is, in his words, "a job for the historians". That the workbooks contain something resembling Cohen's proof does not mean Gödel had it. It shows he was on the path, and there is no credit for work you tell nobody about.

Some points I missed while reading:

  • I found no publication against which the numbers could be checked. Neither the 650 pages nor the Coq files are accessible so far.
  • The report does not say which system the reconstructed proof is for. Gödel's letters speak of type theory, Cohen's result concerns set theory. That is a considerable difference.
  • Which version of Claude was used is not stated anywhere.
  • The dozens of anticipations of later results and the themes nobody has pursued to this day are, for now, statements by von Plato and in part by Claude itself. Whether a result is really absent from the literature is hard to prove, and it is a typical source of error for a language model.
  • Does a repaired proof confirm Gödel's theorem or a new one that resembles it? For about 200 theorems that is the actual historical question.

What I take from it

I build iOS apps and this website with coding agents, Claude Code among them. The workflow in Helsinki looks familiar, only with higher stakes. The model is strong where humans are weak: keeping a lot of material in view at once and connecting passages that are 600 pages apart. The result can only be relied on once there is a check behind it that cannot be talked round. In software development that is tests and compilers, in mathematics a proof assistant.

The ratio is what stands out to me. Of 663 theorems, about 250 failed on the first pass or had to be changed. Without Coq they would have stayed in the analysis, fluently worded and plausible. The post on AI-assisted programming one year later was about the same gap between what looks convincing and what is true.

The headlines lose the fact that the groundwork was done by people. Without the years in which a small group transcribed a forgotten school script sign by sign, Claude would have had nothing to read.

Von Plato wants to turn to Gödel's notebooks on physics and cosmology next. There is a history there too: in 1949 Gödel found a solution of Einstein's field equations that describes a rotating universe in which travel into one's own past is mathematically possible. For the logic workbooks, the specialists' work is still to come: reading what is in the workbook, what the transcription makes of it and what Claude added.

FAQ

Did Claude decipher Gödel's shorthand? No. Jan von Plato and his team transcribed the Gabelsberger shorthand themselves. Claude analysed those transcriptions together with von Plato's books and further scholarly literature.

How many theorems did Claude find in Gödel's workbooks? 663 theorems for which Gödel had noted a proof, a sketch or an idea. The proof assistant Coq verified about 400 as stated, about 200 had to be made more precise or extended, and about 50 were false.

What is Gabelsberger shorthand? A German stenography that Franz Xaver Gabelsberger published in 1834 and that was replaced by the unified German shorthand in 1924. Gödel learned it at school and wrote most of his notes in it.

Did Gödel prove the independence of the axiom of choice before Paul Cohen? That is not established. In letters of 1966 and 1967 Gödel wrote that he had found such a proof for type theory in 1942, but he could no longer reconstruct it. According to von Plato one step was missing in the workbooks: one definition was wrong, and Gödel had noted the right one two years earlier. How much of the version reconstructed with Claude is Gödel's is for historians to settle.

What does Coq check, and what does it not? Coq, now Rocq, checks whether a formally written proof follows from its assumptions. It does not check whether that formal version matches what Gödel meant in his workbooks.

Have the results been published? So far they are known only from a newspaper report. I found no scholarly publication, nor the 650-page analysis or the Coq files.

Sources