Explainer18 min read24 sources
AI and Kurt Gödel's shorthand notebooks: what Claude read, what was checked, what is published
Jan von Plato says Claude, reading his transcriptions of Gödel's shorthand workbooks, listed 663 claims and rebuilt a 1942 argument. What was checked, what is public.
The Helsinki logician Jan von Plato says that Anthropic's Claude, working from his English transcriptions of Kurt Gödel's shorthand workbooks of 1940–42, listed 663 claims in them that Gödel proved, sketched or had an idea for, and rebuilt from Gödel's 1942 notes an almost complete argument for the independence of the axiom of choice, a result Paul Cohen proved in 1963. Gödel “was just one step away,” von Plato told the Guardian, with one wrong definition, and he says the reconstructed proof passed the proof assistant Coq.1 For the 663 claims he gives rough counts from Coq: about 400 held as stated, about 200 after repair, and about 50 were false. He presented the work in a lecture at the University of Vienna on 19 October 2026,3 and it was reported by Der Standard the same day and by the Guardian on 13 November.
Much less is established. Claude did not read Gödel's handwriting: von Plato read the Gabelsberger shorthand, translated it, and said in Vienna that these AI systems cannot read it. As of 21 November 2026 nothing from the work has been published. We found no paper, preprint, transcription, list of the 663 claims or Coq files from von Plato, from his collaborator Josef Urban or from Anthropic, and no source names a Claude model version. The figures are von Plato's own, given in a talk and to journalists, and they differ slightly between tellings. No outside specialist has published an assessment, and Joel David Hamkins of the University of Notre Dame told the Guardian he had not seen the reconstruction.
Below: key terms; what von Plato reports; why Claude worked from transcriptions; a table of who did what; what the 663 and the Coq counts mean; what historians already knew about Gödel and the axiom of choice, and what is new; a timeline; what has been published; and what a record of archive work with AI can note down. Every source was checked on 21 November 2026.
Key terms#
- Gabelsberger shorthand. A German shorthand invented by Franz Xaver Gabelsberger in 1817 and published in 1834, used mostly in southern Germany and Austria until a unified German shorthand replaced it in the 1920s. Gödel learned it at school between 1919 and 1921.9
- Arbeitshefte. Gödel's 16 mathematical workbooks, almost entirely in Gabelsberger.13
- Axiom of choice. The principle that from any collection of non-empty sets one can pick one element from each, even when there are infinitely many sets and no rule for picking.
- Independence. A statement is independent of a set of axioms when the axioms can neither prove it nor disprove it. Showing this takes two halves: a model of the axioms where the statement holds, and one where it fails.
- Forcing. The method Cohen invented in 1963 to build models of set theory in which statements such as the axiom of choice fail.
- Coq, now Rocq. A proof assistant: a program in which proofs are written formally and then machine-checked by a small “kernel”.19 It was renamed the Rocq Prover; the reports use the old name.
What does von Plato say Claude found in Gödel's notebooks?#
Since 2017 von Plato has led a project at the University of Helsinki on 16 notebooks Gödel wrote in the early 1940s, which the Guardian puts at 1,009 pages. Progress on the mathematics was “about a page a week”, the Guardian reports, until September 2026. In August, at a conference in France on AI and theorem proving, computer scientists suggested he try Claude. Back in Helsinki he began “showing Claude his transcriptions of the workbooks, as well as six books he had previously written on Gödel and other secondary scientific literature”.1 Der Standard dates the start to 24 September 2026, when he fed hundreds of pages of those transcriptions into Claude and asked for analyses of Gödel's arguments.2
He describes three results.
An argument on the axiom of choice. His first question was whether the workbooks held a proof that the axiom of choice is independent. “The answer came back: yes, almost,” the Guardian writes. Von Plato: “There's one definition that Gödel got wrong, but he had the right definition two years earlier”.1 According to Der Standard, after a few prompts Claude produced a proof built on the approaches in Gödel's notes and following a different route from Cohen's.2 Von Plato says he ran the reconstructed proof through Coq and it passed.
An explanation of the whole series. “In a couple of days, I got to see a 650-page explanation of the workbooks. It had to be controlled page for page,” he told the Guardian. He added: “The AI can connect all the pieces in a thousand pages. It can see all of Gödel's writings as one piece.”1 His lecture abstract speaks of “some 500 pages” produced “In an interactive process”.3 In the lecture itself he said the explanation had since grown to 1,084 pages, and that the full document covering every claim ran to several thousand. “One question is if this is now humanly possible to read. I don't know.”4
A catalogue of 663 claims. He asked Claude to extract every result for which Gödel “either had a proof, had sketched a proof or had an idea for a proof”. Claude found 663. His Coq counts are discussed below.
He also reports that the workbooks hold, as Claude puts it, “more than a dozen themes in logic and foundations of mathematics” that nobody has taken up in the 80-plus years since.1 In Vienna he said Claude gave him 14 such topics and then 50 pages of elaboration, which he had not yet gone through.4 The lecture went on to Claude's readings of Gödel's notes on quantum mechanics and on a rotating universe; we leave those aside here.
Could Claude read Gödel's Gabelsberger shorthand?#
By von Plato's own account, no: a person deciphered the shorthand, and the model read the result. In the discussion after his lecture, a listener said ChatGPT had been unable to read a family diary written in Gabelsberger. Von Plato replied that these systems cannot read the shorthand, which was what he had expected.4 What Claude read was his transcription and English translation.
That puts a person's reading underneath everything the model produced. Reading Gödel's shorthand is hard even for trained readers. The Helsinki project describes it as “a combination of syntax and an understanding of the content of what Gödel is saying, with a clear dose of guesswork.”9 Every claim Claude listed rests on choices made at that step.
Few people have done that work. Gödel's papers are owned by the Institute for Advanced Study, which deposited them at Princeton University Library in 1985. The finding aid lists 30.75 linear feet of material and says that “a substantial part is in Gabelsberger shorthand”.12 For the edition of Gödel's Collected Works (five volumes, 1986–2003), Cheryl Dawson learned the script from Hermann Landshoff, a retired photographer, and later worked with Tadashi Nagayama. Substantial parts of the notebooks were transcribed and translated, but the editors judged the result not ready for publication.13
Since then two editions have advanced. Eva-Maria Engelen has edited Gödel's philosophical notebooks, the “Max Phil” series, at the Berlin-Brandenburg Academy of Sciences and Humanities, in volumes from 2019 to 2026.20 Von Plato published Can Mathematics Be Proved Consistent?21 (2020), on Gödel's shorthand notes on incompleteness, Chapters from Gödel's Unfinished Book on Foundational Research in Mathematics22 (2022), and, with Maria Hämeen-Anttila, Kurt Gödel: Results on Foundations23 (2023). The project ran as the European Research Council grant GODELIANA10 from September 2018 to August 2024. Its successor, HERITAGE11, began in December 2025 with the stated aim of transcribing and translating into English Gödel's 16 Arbeitshefte, alongside notes by Gerhard Gentzen and Paul Bernays.
Who did what?#
| Step | Who or what, as reported | Source |
|---|---|---|
| Keeping and filming the pages | Princeton University Library holds the papers; the whole collection was microfilmed in 1998. Which images von Plato used is not stated | Finding aid12 |
| Reading the shorthand | Von Plato, who taught himself Gabelsberger | Guardian1 |
| Transcription and English translation | Von Plato | Der Standard2 |
| Background given to the model | Von Plato's six books and other literature | Guardian |
| Explaining and connecting passages | Claude, “In an interactive process” with von Plato | Lecture abstract3 |
| Listing the 663 claims | Claude, at von Plato's request | Guardian |
| Rebuilding the choice argument | Claude, after prompts, from Gödel's notes | Der Standard |
| Modern and formal statements | Claude, as von Plato describes his method: he asks for journal-style proofs and lets Claude handle the proof assistant | Lecture4 |
| Coq checking | Von Plato for the choice proof; the 663 with Josef Urban; no files public | Guardian, Der Standard |
| Page-by-page control of the explanation | Von Plato | Guardian |
| Reading by others | Akihiro Kanamori read the choice argument, von Plato says; Hamkins has not seen it | Lecture, Guardian |
| Historical judgement | Open | Guardian |
Two specialists are named as having seen the work, and neither has published anything about it. In Vienna von Plato said that Kanamori, who did not know the text was “an interactive product of mine and Claude's”, called it “extremely incisive” in both mathematics and history.4 Kanamori is von Plato's co-author on Gödel's set theory,24 so he is not an outside referee. Von Plato also said he wrote to Robert Solovay and received an answer. What Solovay said was not given in the talk as we could hear it. Sandra Müller of TU Wien told Der Standard she was impressed. The article does not say whether she had read the reconstruction.2
What does “663 theorems” count, and what does “about 400 held up in Coq” mean?#
The 663 is the number of items Claude extracted when asked for every result with a proof, a proof sketch or a proof idea in Gödel's notes. In Vienna von Plato called them “663 claims that can be called theorems”.4 So the count mixes finished proofs with sketches and ideas. It counts statements and says nothing about how many are new. For some items, he said, the explanations ran to more than 1,000 pages because Gödel had no proof and one had to be developed.
On the Coq results the two tellings differ slightly:
| Guardian, 13 November | Vienna lecture, 19 October | |
|---|---|---|
| Held as stated | About 400 | 400 |
| Held after repair | About 200 | 200 |
| False | About 50 | 55 |
| Still in progress | Not given | About five to eight |
“Repaired” has a definite meaning in his words: “perhaps the statement of the theorem had to be made more precise, or some assumption had to be added.” On the false ones he added that Gödel sometimes wrote “falsch!” in the margin himself.1
A headline on 16 November summed this up as “About 400 Held Up in Coq”.7 It helps to be exact about what such a check covers. A proof assistant checks that a formal proof proves a formal statement, given the assistant's kernel and axioms. Here each formal statement is a modern reading of an English translation of a shorthand note. Von Plato said in Vienna that “the starting point is always Gödel's own formulation of the result”, that he asks Claude for a proof in “standard mathematics journal article style”, and that Claude handles the interface with the proof assistant, so the user need not read the code.4 Der Standard describes Urban as a specialist in AI-assisted proof verification, with whom von Plato is checking the workbook proofs.2 No source says who compared the formal statements with Gödel's text.
So a Coq success shows that a modern formal statement has a proof the checker accepts. It does not show that the statement is what Gödel wrote, that Gödel had a proof, or who deserves credit. For the 200 repaired items it shows that a nearby, corrected statement holds. A member of the audience in Vienna raised exactly these points: formal statements that drift from the intended problem, and bugs in proof checkers. The same limits apply to Lean, as our guide to checking a Lean proof that comes with a paper explains, and our post on Fermat's Last Theorem in Lean shows them for a machine check of a known theorem. None of the Coq files can be inspected yet.
Did Gödel prove the independence of the axiom of choice?#
Not in anything he published, and his own claim was narrower than some coverage says. The history before 2026 is well documented.
In 1938 Gödel announced, and in 1940 published in full, that the axiom of choice and the continuum hypothesis are consistent with the other axioms of set theory, using his “constructible” sets.1517 That was the first half of independence. The second half, a model where choice fails, came from Cohen, who found forcing in 1963 and received the Fields Medal in Moscow in 1966.16
Alonzo Church, who was to speak on Cohen's work at that congress, asked Gödel whether anything should be credited to him. Gödel's reply of 29 September 1966, as Solomon Feferman, an editor of Gödel's Collected Works, reports it, was that “he [Gödel] only had a proof of the independence of the axiom of constructibility in type theory, which, he believed, could be extended to an independence proof of the axiom of choice.” He added: “The partial result mentioned was never worked out in full detail or put in form for publication.”13 Cohen later gave as one reason for the lack of progress before him the rumour that Gödel had “partially solved the independence problem, at least for AC”.14
The notes themselves were looked at too. Gödel's own index to the Arbeitshefte lists as item 32 “Independence of the axiom of choice; For Zermelo set theory; Systems of representation; Consistency of the existence of lawless sets.” The Collected Works editors gave the transcribed extracts to Robert Solovay and Donald A. Martin. In Feferman's account, “The best that they could come up with was that it looks like a form of forcing, but they were unable to extract a coherent proof of its independence from Gödel's working notes.”13 The Helsinki project's own description, written around 2019, lists “his 1942 proof of independence of the axiom of choice” among results it expected to bring out.8
What is new in 2026 is von Plato's claim, with Claude's reconstruction behind it, that the 1942 argument is complete apart from one definition. As we understood the recording, Gödel builds “possible worlds” out of infinitely many binary sequences, some following a law and some lawless, and the argument ends with a diagonal step of the same shape as his incompleteness proof. The fix replaces a condition stated through definability in a formal language with one about invariance when places are exchanged, which von Plato says Gödel had worked out elsewhere in his notes.4 He also said the notes contain a complete proof, with nothing added, of the independence of the axiom of constructibility, matching what Gödel wrote in 1966. The restored recording's description says von Plato “distinguishes Gödel's original ideas from the additions needed for a modern reconstruction”.5 The setting is type theory, as in Gödel's 1966 statement, and differs from the Zermelo–Fraenkel set theory (ZF) of Cohen's proof. The Proofs and Prompts newsfeed's phrase “independence of the Axiom of Choice to ZF” is looser than that.6
Two details in the coverage are off. The Guardian says Gödel claimed he “had already proved the result”; his 1966 statement claimed less.1 And it reports von Plato's view that Gödel would have won “the 1940 Fields medal”. The International Mathematical Union lists medals in 1936 and then in 1950, with none between.18
Outside comment so far is careful. Hamkins told the Guardian that if the reconstruction stands up “it is remarkable and historically interesting. Although how much is Claude and how much is Gödel is a job for the historians.” He added: “The fact that there are things in there that resemble [Cohen's proof] does not mean he had it. It shows he was on the path. You don't get credit for work you don't tell anyone about.”1 How ideas shared before the evidence came to be credited later is the subject of our post on hypotheses that waited for evidence.
When did each step happen?#
| When | What | Source |
|---|---|---|
| 1938–40 | Gödel shows choice and the continuum hypothesis consistent with set theory | Kanamori 2007, MacTutor |
| 1940–42 | Gödel writes the workbooks in Princeton | Lecture abstract3 |
| 1942 | Gödel's notes on the independence of choice | Helsinki project8 |
| 1963 | Cohen invents forcing and proves independence | MacTutor16 |
| 1966 | Cohen's Fields Medal; Gödel's statement of 29 September | IMU, Feferman |
| 1980s–2003 | During the Collected Works edition, Solovay and Martin study the extracts and see “a form of forcing” | Feferman13 |
| 2017; 2018 | Von Plato starts the project; ERC funding from September 2018 | Guardian, CORDIS |
| August 2026 | Suggestion to try Claude, at a conference in France | Guardian1 |
| 24 September 2026 | Transcriptions first given to Claude | Der Standard2 |
| 19 October 2026 | Der Standard report; Vienna Circle Lecture at 19:00 | Der Standard, poster3 |
| 20 October 2026 | Restored recording posted | ai4reason5 |
| 13 November 2026 | Guardian report | Guardian |
| 16–17 November 2026 | Follow-up coverage and the Proofs and Prompts entry | Newsfeed6 |
The newsfeed dates the lecture to 20 October.6 The poster and the university's livestream put it on Monday 19 October, starting at 7 pm Vienna time; the stream ran from 17:16 to 18:11 UTC.34 The restored version appeared on 20 October.
What has been published, and what has only been said?#
Published before 2026 are von Plato's books on Gödel's notes, the Collected Works and the volumes of the philosophical notebooks. The 2025 article by Kanamori and von Plato on Gödel's first proof of the consistency of choice, based on his notebooks of 1935, also predates the AI work.24
What is public from the 2026 work: two recordings of the lecture, its abstract, and interviews. The restored recording notes that “The camera started after the lecture had begun, so the opening portion is missing.”5 The livestream also starts mid-talk. Our account of the talk rests on its automatic English captions, which mishear some words, so we quote it sparingly.
What is not public, as of 21 November 2026: the transcriptions given to Claude, the prompts, the explanation, the list of 663 claims, the Coq files, the 30-page article on the choice argument that von Plato mentioned, Solovay's reply, and the 14 themes. Anthropic has published nothing on the project that we could find, and nobody has named a model version. In the discussion von Plato said his Claude Max plan would not cover the costlier checks; a plan name does not identify a model.
Several numbers differ between sources:
- Pages. The 2026 reports say 1,009 pages from 1940–42. The project's own page counts the 16 Arbeitshefte at “altogether some twelve hundred shorthand pages”, written from the mid-1930s to the mid-1940s.8 We found nothing that reconciles the two.
- The explanation. About 500 pages in the abstract, 650 in the Guardian, 1,084 in the lecture.
- False claims. About 50 in the Guardian, 55 in the lecture.
- Start. Claude was suggested in August, and Der Standard dates the start to 24 September.
None of this makes the work wrong. It means that what readers have is reports at different moments, without the underlying files.
What this means for a research record#
Archive work with an AI model has more links in its chain than most research, and each one can be written down at the time:
- The originals. Which pages, by collection, box, folder or microfilm reel, and which images or scans were used.
- The transcription. Who read the shorthand, which version of the transcription and translation went to the model, and on what date.
- The model. The product and model version as the provider showed it, the plan, the dates, and what it was asked to do, with the prompts kept.
- The output. What came back, in which version, and which parts a person checked against the original images, page by page.
- The formal check. Who wrote each formal statement, how it compares with the transcription, which proof assistant version ran, and where the files are.
- Outside readers. Who read the result, what they read, and how they are related to the project.
A signed record of these points would not make any reading of Gödel right. It would give historians the material for the question Hamkins raised: which reading the model worked from, what it added, and which statement Coq checked. On Papex, an AI app can prepare a private draft of such a record, and a person reviews and signs what becomes public (how AI apps work with Papex).
A note on interests: this post reports on work done with Anthropic's Claude, and it was drafted with Claude. Readers should weigh it with that in mind and check the sources below.
Sources#
- Alex Bellos, ‘Ahead of his time’: AI helps unlock discoveries hidden in notebooks of man who transformed maths (The Guardian, 13 November 2026), checked 21 November 2026 a b c d e f g h i j
- Reinhard Kleindl, Rätsel um verlorenen Gödel-Beweis gelöst (Der Standard, 19 October 2026), in German, checked 21 November 2026 a b c d e f
- Institut Wiener Kreis, 34th Vienna Circle Lecture: Jan von Plato, Understanding Gödel in an Age of Machine Intelligence (announcement, 19 October 2026), checked 21 November 2026 a b c d e f
- Uni Wien live, Understanding Gödel in an Age of Machine Intelligence (livestream recording and description, 19 October 2026), read through its automatic English captions, checked 21 November 2026 a b c d e f g h i
- ai4reason, Jan von Plato: Understanding Gödel in the Age of Machine Intelligence (restored recording and description, 20 October 2026), checked 21 November 2026 a b c
- Proofs and Prompts, Newsfeed, entry of 17 November 2026, checked 21 November 2026 a b c
- Claude Found 663 Gödel Theorems; About 400 Held Up in Coq (FourWeekMBA, 16 November 2026), checked 21 November 2026 ↩
- Gödel Enigma, Project description (University of Helsinki), checked 21 November 2026 a b c
- Gödel Enigma, Gabelsberger stenography (University of Helsinki), checked 21 November 2026 a b
- The Gödel Enigma: Unveiling a Hidden Logical Heritage, GODELIANA, grant 787758 (CORDIS, European Commission), checked 21 November 2026 ↩
- The hidden heritage of modern logic, HERITAGE, grant 101198605 (CORDIS, European Commission), checked 21 November 2026 ↩
- Kurt Gödel Papers, C0282, finding aid (Princeton University Library), read through the Internet Archive copy of 10 August 2026, checked 21 November 2026 a b
- Solomon Feferman, The Gödel Editorial Project: A synopsis (Bulletin of Symbolic Logic 11, 2005; author's copy), checked 21 November 2026 a b c d e
- Akihiro Kanamori, Cohen and set theory (Bulletin of Symbolic Logic 14, 2008; author's copy), checked 21 November 2026 ↩
- Akihiro Kanamori, Gödel and set theory (Bulletin of Symbolic Logic 13, 2007), abstract and references, checked 21 November 2026 ↩
- Paul Cohen, biography (MacTutor History of Mathematics, University of St Andrews), checked 21 November 2026 a b
- Kurt Gödel, biography (MacTutor History of Mathematics, University of St Andrews), checked 21 November 2026 ↩
- Fields Medal, list of laureates (International Mathematical Union), checked 21 November 2026 ↩
- About the Rocq Prover, formerly Coq (Rocq team), checked 21 November 2026 ↩
- Kurt Gödel, Philosophische Notizbücher / Philosophical Notebooks, edited by Eva-Maria Engelen (De Gruyter Brill, series page), checked 21 November 2026 ↩
- Jan von Plato, Can Mathematics Be Proved Consistent? Gödel's Shorthand Notes & Lectures on Incompleteness (Springer, 2020), bibliographic record, checked 21 November 2026 ↩
- Jan von Plato, Chapters from Gödel's Unfinished Book on Foundational Research in Mathematics (Springer, 2022), bibliographic record, checked 21 November 2026 ↩
- Maria Hämeen-Anttila and Jan von Plato, Kurt Gödel: Results on Foundations (Springer, 2023), bibliographic record, checked 21 November 2026 ↩
- Akihiro Kanamori and Jan von Plato, Gödel's First Proof of the Consistency of the Axiom of Choice (History and Philosophy of Logic, 17 September 2025), abstract, checked 21 November 2026 a b