Skip to content

Claude helped decode Gödel's notebooks and found 663 theorems

A robot at an archive desk reads a stack of old notebooks as their pages turn into glowing math symbols and proof trees Models

Claude helped a Finnish logician work through Kurt Gödel's notebooks from the early 1940s. The notes run to 1,009 pages, and scholars had been reading them at about a page a week. The model listed 663 theorems, proofs and proof ideas in them. Among them is an almost complete proof that Gödel claimed to have found in the 1940s but never published, The Guardian reported on November 13.

The researcher is Jan von Plato of the University of Helsinki, who has studied Gödel's unpublished papers for years. He checked the results in Coq, a proof assistant that verifies mathematics line by line. Most of the 663 items held up, but not all of them, and the hardest questions are still open.

The story starts with the notebooks themselves.

What is in Gödel's notebooks

Kurt Gödel is best known for his incompleteness theorems of 1931. They showed that any consistent formal system strong enough for basic arithmetic contains true statements it cannot prove. In 1940 he left Austria for Princeton, where he worked at the Institute for Advanced Study and became a close friend of Albert Einstein.

In the early 1940s Gödel filled 16 notebooks, which he called workbooks, with mathematics and logic. They run to 1,009 pages, mostly written in Gabelsberger, a German shorthand that was already going out of use when Gödel learned it. Few people alive can read it.

Von Plato has led a project to study the workbooks at Helsinki since 2017, backed by a European Research Council grant reported at €2.5 million. Progress was slow. In his own words, the team needed about a year to read around 50 pages. At that pace, a full reading would have taken another decade or two.

That is why von Plato tried an AI model.

How von Plato used Claude

The suggestion came in August 2026, at a conference in France on AI and theorem proving. Computer scientists there advised von Plato to try Claude, and he started in September.

Claude did not read Gödel's handwriting. Von Plato gave it the text his team had already prepared:

  • his transcriptions of the workbooks;
  • six books he has written about Gödel;
  • other secondary literature on Gödel's work.

Within a couple of days, Claude produced a 650-page explanation of the workbooks. Von Plato told The Guardian he was "flabbergasted" by the result. He called the job close to impossible for humans, who read one page at a time and forget earlier ones. He also said he checked the digest page by page.

Reports do not say which Claude model he used or whether Anthropic was involved. The Guardian did not report the cost either.

The first question von Plato put to Claude was about a famous gap in Gödel's record. The second was broader: list every theorem in the workbooks.

The 663 theorems and the Coq check

Von Plato asked Claude to extract every result for which Gödel had a proof, a sketch of one or just an idea for one. It returned 663 items. That number covers finished arguments, rough outlines and false starts together, so it is not a count of 663 new proofs.

Von Plato then tested the statements in Coq, which has since been renamed Rocq. The reported results:

Result in CoqItems (approx.)What it means
Verified as stated400the theorem holds exactly as Claude extracted it
Repaired200it holds after the statement was made more precise or an assumption was added
False50the statement is wrong
Not accounted forabout 13reports do not explain the rest of the 663

The false results do not surprise von Plato. Gödel sometimes marked his own notes "falsch!" (wrong) or "schlecht!" (bad) when he reread them.

According to von Plato, dozens of the correct results were not published by anyone until the 1970s and 1980s. Claude also flagged more than a dozen logic themes in the workbooks that nobody has followed up in over 80 years. That last point is Claude's reading, and nobody has checked it yet.

The most interesting item on the list is the one von Plato asked about first.

The axiom of choice proof Gödel almost had

The axiom of choice says that, given any collection of non-empty sets, you can pick one element from each. Mathematicians rely on it everywhere, but for decades nobody knew whether it follows from the other axioms of standard set theory.

Gödel solved half the problem. In 1938 he showed that the axiom of choice cannot be disproved from the other axioms. The other half, that it cannot be proved from them either, came from Paul Cohen in 1963. Cohen invented a technique called forcing and won the Fields Medal for it in 1966.

Gödel later said he had proved the independence of the axiom of choice himself in the 1940s. When colleagues pressed him for details, he could not rebuild the proof from his notes.

Von Plato asked Claude whether the workbooks held such a proof. The answer was almost. Claude reconstructed an argument that was complete except for one definition, which Gödel had gotten wrong. Two years earlier, in the same notebooks, Gödel had used the correct one.

"He was just one step away," von Plato told The Guardian. He believes a few more months of work would have given Gödel the result two decades before Cohen. The reconstructed proof, with the definition fixed, passes Coq.

Not every logician is ready to rewrite the history books.

What other logicians say

Joel David Hamkins is a set theorist at the University of Notre Dame. He told The Guardian he had not seen Claude's reconstruction. If it stands up, he said, it would be remarkable and historically interesting. He added that a proof resembling Cohen's shows Gödel was on the path, not that he had the result. "You don't get credit for work you don't tell anyone about," he said.

Hamkins also raised a harder question: how much of the reconstructed proof is Gödel's and how much is Claude's repair work. He called that a job for historians.

A few other limits apply to the whole project:

  • One reviewer so far. Von Plato checked the 650-page digest himself. Outside logicians have not reviewed it in detail.
  • No public data yet. No paper, Coq files or full theorem list have been released, so others cannot rerun the check.
  • Coq checks statements, not history. A passing proof shows the mathematics is correct. It cannot show what Gödel knew or meant.

Next, von Plato plans to work through the repaired and false items and the untouched logic themes. After the logic work, he wants to apply the same method to Gödel's notebooks on physics and cosmology.

The Gödel project also fits a wider pattern in AI and mathematics this autumn.

How this compares with other AI math results

AI has been showing up in several kinds of mathematical work in recent months. The KLS conjecture proof used AI agents to explore about a hundred approaches, while the key idea came from the human authors. Axiom Math used agents to formalize the Stacks Project in Lean and found errors along the way. OpenAI, Anthropic and the Simons Foundation also put $15 million into tools for verifying AI proofs.

The Gödel case is different. Claude did not attack an open problem. It read a huge archive faster than any team could and turned it into a list of claims that a proof checker could test. Archive work like this may be one of the most useful near-term jobs for language models in science, because the output can be checked.

That leaves the question in the headline of most reports.

Did Claude uncover Gödel's lost theorems

Partly. Claude found and organized hundreds of results in notes that humans had barely started reading. About 400 of them pass a formal check as written. That alone would have taken von Plato's team many years. The near-proof of the axiom of choice's independence is the most striking find, because it backs up a claim Gödel made and could never support.

The bigger claims need more evidence. The 663 items are candidates, about 50 are false and the reconstructions have not been reviewed by other logicians. Until von Plato publishes the digest and Coq files, the safe reading is that Claude made an archive searchable and testable. Whether Gödel really had Cohen's result is, as Hamkins put it, a question for historians.

FAQ

What did Claude find in Gödel's notebooks?
Claude listed 663 theorems, proofs, proof sketches and ideas in Gödel's 16 workbooks from the early 1940s. In Coq, about 400 were correct as stated, about 200 needed repairs and about 50 were false. It also reconstructed an almost complete proof of the independence of the axiom of choice.

Did Claude read Gödel's shorthand?
No. Jan von Plato's team had already transcribed the workbooks from Gabelsberger shorthand. Claude worked from those transcriptions, plus von Plato's six books on Gödel and other literature.

Did Gödel prove the independence of the axiom of choice before Paul Cohen?
Not quite. According to von Plato, the workbooks contain a proof that is one definition short. Gödel had the right definition two years earlier, and the repaired proof passes Coq. Other logicians have not yet reviewed it, and Gödel never published it.

Which Claude model was used?
Reports do not say. The Guardian did not name the model or say whether Anthropic worked with von Plato directly.

What is Coq?
Coq, now called Rocq, is a proof assistant. Mathematicians write a statement and its proof in its formal language, and the software checks every step.

Sources

Portrait of Felix Laurent, author at thehype.news

Felix Laurent

News Writer at thehype. Felix is a journalist with experience in marketing. He covers artificial intelligence and emerging technologies, focusing on the ideas, innovations, and trends shaping the future of tech.

More articles by Felix Laurent →
On air · radio.thehype.news ↗ AI news radio — 24/7