Vienna morning edition — November 14, 2026. An original Guardian investigation published November 13 at 11:38 a.m. EST describes how researchers used Claude to help analyze 1,009 pages of notebooks written by logician Kurt Gödel. Some notes were in an obsolete shorthand. The project demonstrates a striking ability to connect material across long documents—but it also shows why generated explanations must be checked independently.
The research, not just the headline
The reporting follows Jan von Plato’s University of Helsinki-led project. Researchers had spent years deciphering Gödel’s workbooks; the AI-assisted pass produced a broad digest quickly. According to von Plato, the model identified 663 candidate theorem statements or ideas. When they were checked using the Coq proof assistant, around 400 were verified as written, approximately 200 needed repair or more precise assumptions, and roughly 50 were false.
Those numbers are reported by the researcher and rounded in the original article; they are not a general benchmark for Claude’s mathematical accuracy. The investigation also describes a possible connection with later independence results in set theory. Experts cautioned that finding a promising argument in old notes is not equivalent to proving that the original author completed or published the later result.
What developers can learn
Long-context synthesis is useful when information is fragmented, hard to read or scattered across many files. However, a plausible digest is not the same as a checked theorem, reliable medical conclusion or lawful contract interpretation. A good product should preserve the path from a generated claim back to its source and validation result.
Practical experiment: an evidence-to-verification pipeline
- Choose a small collection of public technical papers or internal documents you have permission to process.
- Ask a model to extract atomic claims with page and paragraph references, keeping interpretation separate from quotation.
- Run deterministic checks for numbers, formulas or code where possible; send uncertain claims to an expert reviewer.
- Label each result verified, repaired, rejected or not checked. Never let unchecked claims enter the verified bucket.
- Measure time saved, missing citations and false claims over the full workflow—not only model output speed.
Sources and limits
Original reporting: Alex Bellos, The Guardian, November 13, 11:38 a.m. EST (later updated at 1:38 p.m. EST). The researcher descriptions are not presented here as independently replicated experiments. This INGOAMPT analysis emphasizes an implementation lesson; see other AI research and developer coverage. The illustration is original abstract art, not a historical notebook photograph.
