Kurt Gödel filled 16 workbooks with mathematical notes after settling in Princeton, New Jersey, in the early 1940s. Much of the writing used an old German shorthand. The notebooks could take researchers a week to work through one page at a time. In September, after the University of Helsinki’s Jan von Plato supplied transcriptions and other research material to Anthropic’s Claude, the model produced a 650-page analysis in days, the Institute for Advanced Study says.
That speed let von Plato ask a question mathematicians had been unable to settle from the papers. Had Gödel come close to a result that Paul Cohen would make public two decades later? The team also checked hundreds of other proposed results from the notebooks. Some could be proved as written. A substantial number needed changes, and some were false.
The years of work before Claude

Gödel’s workbooks were written partly in Gabelsberger, a shorthand system once common among German-speaking academics. Getting the letters and symbols into a readable form was a specialist’s job. The University of Helsinki’s project description documents the team’s efforts to transcribe and translate his papers, including the workbooks and a separate set of notebooks on foundations. An English edition of those other notebooks shows what the earlier work produced: it presents four notebooks with 126 numbered theorems.
The 16 workbooks in the new account run to 1,009 pages. Von Plato gave Claude transcriptions he had already made, material from his own books on Gödel, and other literature. Specialists first made the material readable. Claude analyzed the resulting text across hundreds of pages, helping connect passages a reader might otherwise study months apart.
What the proof checker could establish
Von Plato told the reporter who interviewed him that the model extracted 663 candidate results, including statements Gödel had proved, sketched or considered. About 400 held as stated when checked. Roughly 200 could be repaired, for instance by stating a claim more precisely or adding an assumption. About 50 were false. Those rounded figures describe different outcomes in one project. They give no general accuracy rate for the model.
The checks used Coq, the software now called the Rocq Prover. It verifies a proof expressed in a formal language by checking whether each step follows from the definitions, assumptions and rules entered into it. Once the claim has been written in that formal language, a reader can have the software check the same proof again. Deciding whether the definitions in that proof match what Gödel used still requires comparing the formal version with the historical manuscript.
How close did Gödel get to Cohen’s result?

The axiom of choice says, roughly, that it is possible to select one member from each set in a collection of nonempty sets, even when the collection is infinite. The independence question asks whether the other standard axioms of set theory can settle that statement. Cohen’s published work and his 1966 Fields Medal are part of the public record.
There are two sides to independence. Gödel had already shown that adding the axiom of choice to standard set theory need not introduce a contradiction. The missing direction was whether the other axioms could coexist with its failure, the result Cohen later published. The Stanford Encyclopedia of Philosophy’s account sets out Gödel’s published work.
Von Plato said the workbooks contained much of an argument for that missing direction, with an incorrect definition at a decisive point. He said Gödel had used the correct definition two years earlier, and that a reconstructed proof passed Coq. The check establishes the reconstructed argument’s validity in the formal setting used. The two manuscript passages leave historians with the question of how much of that argument Gödel himself put together. The Institute’s archive guide identifies the Gödel papers as being on deposit at Princeton University, and von Plato’s team plans to turn to the physics and cosmology notebooks next.






