Claude deciphers Gödel's notebooks: 663 unknown theorems extracted from 1,009 pages of shorthand
🔎 Nine years at one page per week, then three days
For nine years, roughly the only researcher in the world able to read Kurt Gödel's notebooks moved at a pace of one page per week. In September 2026, he handed the corpus over to Claude. A few days later: a 650-page digest, 663 catalogued theorems, and a near-proof of something the Viennese logician had been claiming since 1966.
The news, revealed by The Guardian on November 13, 2026, is no mere media buzz. The figures have been cross-checked, the theorems sifted through the Coq proof assistant, and the man at the center of it all — Jan von Plato, University of Helsinki — admits to having been "flabbergasted".
Why it matters: this is one of the first cases, at this scale, where an LLM deciphers a scientific corpus that is illegible to nearly all of humankind — and where every claim it produces is formally verified. This isn't text generation. This is research.
The Essentials
- The corpus: 16 of Gödel's notebooks from the 1940s — 1,009 dense pages of mathematics and logic, written in the obsolete Gabelsberger shorthand — of which only about six people in the world can still read.
- The turning point: after years at ~1 page/week, von Plato gives Claude his transcriptions, his six books on Gödel, and the scientific literature. Within days: a complete 650-page digest, verified page by page.
- The tally: 663 theorems extracted (full proof, sketch, or idea). Fed into Coq: ~400 verified as-is, ~200 repaired, ~50 false.
- The showstopper: Claude reconstructed a proof of the independence of the axiom of choice, one definition short of complete. Once repaired, it passes Coq. Gödel's claim (1966) to have reached this result back in the 1940s — before Cohen's Fields Medal — now stands credited.
- The lesson: AI replaces neither the expert nor formal verification. It amplifies the one and automates the other.
Recommended Tools
| Tool | Main use | Price (November 2026) | Ideal for |
|---|---|---|---|
| Claude (Anthropic) | Corpus deciphering, document digestion, theorem extraction | From $20/month (check on anthropic.com) | Researchers, historians of science, philologists |
| Coq | Formal verification of statements and proofs | Free, open source | Logicians, mathematicians |
| ChatGPT (GPT-5.5) | General-purpose reasoning, comparison on its own corpus | ≈ $20/month for ChatGPT Plus (check on openai.com) | Teams that want to compare multiple models |
| Gemini 3 Pro Deep Think | Analysis of very long documents, mathematical reasoning | Via Google AI subscription (check on gemini.google.com) | Large corpora, work on long contexts |
A word about this table: for this type of project, the winning architecture is not "an LLM" but "an LLM + a verifier." The LLM proposes, Coq disposes. It's this duo that makes the experiment reproducible — and credible.
One man, 1,009 pages and six readers in the world
Direct answer: without Jan von Plato, this story doesn't exist. AI didn't "decipher" Gödel on its own — it multiplied the last expert capable of reading him.
Since 2017, this Finnish logician at the University of Helsinki has been leading a project funded by the European Research Council to the tune of €2.5 million, devoted to Gödel's 16 workbooks from the 1940s. 1,009 pages dense with mathematics and logic, written for the most part in Gabelsberger shorthand — a German system from the 19th century, now dead, of which von Plato is one of only six people in the world able to decipher it.
The pace? About one page per week, until September 2026. At that rate, getting through the entire corpus would have taken an entire career. And time was running out: experts capable of reading this shorthand are not legion, nor are they eternal.
Because the stakes go far beyond biographical anecdote. Gödel is the author of the 1931 incompleteness theorems, the result that shook the foundations of mathematics — and which would inspire Turing in 1936 to lay the foundations of modern computing, as Archyde reminds us. His drafts from the 1940s are not lecture notes: they are mines of never-published results, written by one of the most rigorous minds of the century.
How Claude Deciphered Gabelsberger Shorthand
Direct answer: by feeding the model all the available expert context — not with a magic prompt.
The trigger was a meeting. In August 2026, at a conference in France on AI and theorem proving, computer scientists suggested von Plato try Claude. A few days later, the logician gave the model his own transcriptions, his six books on Gödel, and the scientific literature in the field — Dataist.ai details this method.
The result: within days, Claude produced a complete 650-page digest of the notebooks, which von Plato verified page by page. A task that, in his view, no human could have carried out.
What interests me here is the method — and there's nothing exotic about it. Von Plato did, without naming it, exactly what we describe in our article on context files: CLAUDE.md, AGENTS.md and beyond: giving the model a reference corpus, precise vocabulary, and examples of validated transcriptions. The model didn't learn Gabelsberger from scratch — it anchored the shorthand in the mathematical context and in the expert's transcription style.
That's the difference between "decipher this for me" and "here's ten years of my work, carry on." The second formulation works. The first doesn't.
One point of caution, though: the digest was reviewed page by page by the human. The AI produced, the expert validated. The order of operations matters, and it must never be reversed.
663 theorems, three fates: Coq's verdict
Direct answer: of the 663 theorems extracted from the notebooks, roughly 400 pass Coq's formal verification as-is. That's a remarkable rate — and the ~50 false ones prove that verification remains indispensable.
The numbers, confirmed by MLQ.ai:
| Verdict in Coq | Count | Meaning |
|---|---|---|
| Verified as-is | ~400 | The statement and proof pass mechanical verification without modification |
| Repaired | ~200 | Statement tightened or hypothesis added: the result holds, the formulation doesn't |
| False | ~50 | Flawed proof or invalid statement — discarded |
Let's read this table honestly. About 7.5% of the theorems proposed by Claude are false. If a PhD student turned that in, they'd get a scolding. But let's put things in perspective: these 663 theorems were extracted from pages that only six people in the world know how to read, in a matter of days, with over 90% valid results after light repair. The cost/benefit ratio isn't even up for debate.
And above all: the ~50 false ones never came close to making it into the scientific literature. Coq caught them. That's exactly the role we want formal verification to play — a mechanical safeguard, immune to the charm of well-written prose.
To reproduce this pipeline yourself, Coq installs in two commands:
opam install coq
coqc mon_theoreme.v
The hard part isn't the tool. It's translating an informal mathematical statement into a verifiable specification — and there, the expert stays at the controls.
The showstopper: the independence of the axiom of choice
Direct answer: the gem of the notebooks is a near-complete proof of the independence of the axiom of choice — the result that won Paul Cohen the Fields Medal. Claude reconstructed it, one definition short of complete; once repaired, it passes Coq.
For non-logicians: the axiom of choice is one of the most debated principles in modern mathematics. Proving its independence — that is, showing that it can be added or removed without making the theory inconsistent — earned Paul Cohen his Fields Medal in 1966.
Now, according to von Plato's reconstruction, Gödel claimed in 1966 to have reached this result as early as the 1940s, in those famous notebooks. Until now, no one had been able to verify it. Claude rediscovered the proof, missing just one definition; once repaired by the human, it passes formal verification. Gödel's claim is thus credited — a quarter of a century before the official recognition of the result.
And that's not all. The notebooks anticipate results that would not be published until the 1970s–1980s, and contain 14 topics in logic never explored to this day. Fourteen research programs, in sum, offered to the community — signed by a man who died in 1978, transcribed by a machine, validated by an expert.
Von Plato's reaction, as reported by the Guardian: "I was flabbergasted." In other words: "I was stunned." When the world's foremost expert on the subject admits to being floored, the news item becomes a fact of science.
AI, the New Microscope of the History of Science
Direct answer: this affair is the best argument to date for AI as a research tool in the history of science — provided you accept the "expert + machine + verifier" architecture.
What happened in Helsinki is reproducible elsewhere, and that's what really excites me. Vast masses of scientific archives lie dormant in dead scripts: stenography, shorthand, handwritten correspondence. Every corpus has its potential von Plato — a rare expert, often aging, who moves at one page per week. AI's leverage on this bottleneck is brutal: 650 pages processed in a few days, where it would have taken roughly twelve years at one page per week.
My opinion, straight up: this is more important than 90% of the AI use cases we're being sold. Here, we're talking about unknown mathematical results, entire research programs (14 themes!), intellectual heritage returned to the community. Not generating slides or LinkedIn posts.
Three conditions, however, for this to hold up:
- An available expert, to provide context and validate. Without von Plato, no reliable digest.
- Formal verification when the field allows it (Coq, Lean...). Without Coq, the ~50 false theorems would have gone unnoticed.
- Total transparency about what is verified and what is not. The 650-page digest was reread page by page — that's not negotiable, it's the standard to enforce.
Why Claude — and should you trust the model choice?
Direct answer: von Plato hasn't published any comparative benchmark, but Claude's profile fits the work required — and the competition has nothing to envy it on paper.
A refresher on the landscape, based on the agentic benchmarks from June 2025:
| Model | Vendor | Agentic score |
|---|---|---|
| GPT-5.5 | OpenAI | 98.2 |
| Gemini 3 Pro Deep Think | 95.4 | |
| Claude Opus 4.7 (Adaptive) | Anthropic | 94.3 |
| GPT-5.4 Pro | OpenAI | 91.8 |
| Claude Opus 4.6 | Anthropic | 84.7 |
What von Plato needed: following transcription instructions to the letter, digesting a massive corpus, inventing nothing beyond the provided context. That's disciplined agent work, not unbridled creativity. Claude Opus 4.7 (Adaptive) is tailor-made for this task profile — and experience seems to confirm it.
That said, let's be honest: no one has tested GPT-5.5 or Gemini 3 Pro Deep Think on Gabelsberger. The choice of Claude owes as much to the August 2026 meeting — computer scientists who recommended it — as to any ranking. For a similar project, start from your corpus and compare for yourself: our Claude vs ChatGPT comparison and the selection of best LLMs for coding (theorem proving is a direct cousin of code) are two good starting points.
For a broader framework, read Claude, GPT, Gemini, Llama: which model should you choose in 2026? and the unfiltered take of Claude 4 vs GPT-5 vs Gemini 3: the honest comparison no one does.
Signal to watch: Anthropic doesn't intend to rest on its laurels with this showcase — the vendor just announced Fable 5.2 and Mythos 5.2, presented as 30% faster and 20% cheaper. The next "AI + archives" deal will probably come with costs divided even further.
❌ Common Mistakes
Mistake 1: Confusing the digest with the proof
The 650-page digest is not a mathematical publication. It's an extraordinarily detailed map of a territory — not the territory itself. What counts are the statements verified in Coq. Solution: never cite a "Claude theorem" without specifying its verification status (verified, repaired, false).
Mistake 2: Skipping formal verification "because it sounds right"
~50 out of 663 theorems are false. A well-written text, with plausible notation, persuades — that's precisely the danger of LLMs. Solution: whenever the domain allows it, everything goes through a proof assistant. When it doesn't, expert page-by-page review is the bare minimum — that's what von Plato did, and it's no coincidence.
Mistake 3: Throwing an illegible corpus at the AI without context
Without the transcriptions, the six books, and von Plato's literature, Claude would have produced nothing but a jumble of stenographic hallucinations. Solution: build the context first — validated examples, glossary, reference corpus — using the context file method described above. Context is 80% of the result.
Mistake 4: Concluding that the expert has become optional
Von Plato is one of only six people in the world able to read Gabelsberger, and the author of six books on Gödel. AI multiplied that expertise; it did not replace it. Solution: budget human validation time as much as machine time. It's what guarantees the scientific value of the final result.
❓ Frequently Asked Questions
Did Claude "prove" new theorems?
No, and that's the key point. Claude extracted and reconstructed results contained in the notebooks: roughly 400 verified as-is by Coq, ~200 after repair, ~50 false. The proof of the independence of the axiom of choice required human repair before it could pass verification. The AI proposes, the machine verifies, the expert has the final say.
What is Gabelsberger shorthand?
A German shorthand system invented in the 19th century by Franz Xaver Gabelsberger, written in cursive and highly abbreviated. Long the standard in Germany and Austria, it fell into disuse in the 20th century. Gödel used it in his notebooks, and only about six people in the world can still read it fluently — including Jan von Plato.
Why are the notebooks from the 1940s so important?
Gödel had just upended the foundations of mathematics with his incompleteness theorems (1931). The notebooks show that he anticipated results not published until the 1970s–1980s, including the independence of the axiom of choice (Cohen's Fields Medal), and contain 14 logic themes never explored. A decade ahead of his field, sealed away until now.
Can this kind of project be replicated?
Yes, subject to three conditions: a corpus expert to provide context and validate the outputs, high-quality digitization of the documents, and a formal verification pipeline where the field lends itself to it. The marginal cost is low compared to the €2.5 million ERC project — the LLM processed nine years of human work in a few days.
Which model should you choose for a similar project?
Prioritize context fidelity and instruction following, not creativity. Claude Opus 4.7 (Adaptive) has proven itself on this use case, GPT-5.5 dominates agentic benchmarks, and Gemini 3 Pro Deep Think excels on long documents. Test on a sample of your corpus before committing — and keep the formal verifier.
✅ Conclusion
In just a few days, Claude accomplished what nine years of solitary toil had failed to achieve: opening the 1,009 pages of Gödel to the world — 663 theorems, a near-proof worthy of a Fields Medal, 14 novel research avenues, all under the dual oversight of an expert and a formal verifier. If you take away only one thing from this story, remember the architecture to replicate: expert + LLM + mechanical proof. To choose the model capable of carrying this duo into your field, start with our Claude vs ChatGPT comparison.