Case Studies

AI Agent Makes Gödel’s Theorems Searchable

AI is helping transcribe, index, and formally organize hundreds of proven theorems in Kurt Gödel’s unpublished manuscripts. Its practical value lies not in autonomous mathematical discovery, but in turning difficult archival material into a searchable and verifiable research resource.

Font size

The Guardian reports that AI has helped organize hundreds of proven theorems in Gödel’s manuscripts. The work involves a clear division of labor: AI handles transcription, retrieval, and initial screening, while mathematicians assess context and verify proofs.

AI Agent Makes Gödel’s Theorems Searchable

This work is easily mistaken for another case of “AI solving mathematics.” What actually happened was more specific. A research team used AI to interpret and formally organize Gödel’s manuscripts. According to The Guardian, the papers contain hundreds of theorems that had already been proved but had not yet been fully organized or published.

The key question was not whether a machine could guess new theorems, but whether AI could connect manual proofreading, retrieval, and verification into a single workflow.

This Was Not an Exercise in AI-Generated Theorems

The distinction matters. These results did not come from asking a large language model to invent a new proof from scratch.

Gödel’s original material already existed. The target was a body of manuscripts that had never been fully organized: some notation was difficult to recognize, definitions depended on context, and arguments were scattered across pages. The team used AI to assist with transcription, retrieval, and formal organization. Mathematicians still had to determine whether each theorem was valid.

This differs from the process of tackling a public mathematical conjecture. A public conjecture usually provides a single proposition. Gödel’s manuscripts are closer to an incomplete research archive. A passage may contain a complete proof, or it may hold only a draft, a partial argument, or intermediate steps that make sense only alongside other manuscripts.

AI was therefore not a “stand-in for a mathematician.” It functioned more like a high-recall first-pass filter.

The system searched large volumes of manuscript pages for passages that might correspond to a proven proposition, normalized notation, extracted structure, and passed candidate results to experts. AI supplied leads, not conclusions.

This boundary is important. When the source material already contains a proof, the task is primarily to interpret an existing manuscript. Producing a plausible-looking proof is a different problem and should not be conflated with transcription and verification.

Why Hundreds of Theorems Justify the Investment

The project team had practical reasons for committing computing power and researchers. For hundreds of theorems, the bottleneck is often not reasoning but reading, checking, and cataloging.

Before handwritten material can enter a research workflow, it typically passes through several manual stages:

  • Recognizing formulas and custom notation;
  • Determining which statements are conjectures and which have proofs;
  • Restoring definitions and context;
  • Checking references to classical results in each proof;
  • Numbering each proposition and recording its place in the research sequence.

These tasks are repetitive and time-consuming, and they are often postponed because experts have limited time. AI can perform an initial round of classification and transcription, narrowing the range of pages that experts must inspect line by line.

A common research strategy is to expand the candidate pool first, then increase the precision of human review. Rather than asking people to read several hundred pages line by line, a model can identify passages that may be worth examining.

High recall, however, solves only half the problem. If AI omits a crucial definition, the formal chain that follows may still break. The system must also record each theorem’s premises and dependencies so experts can trace them backward.

The Real Bottleneck Is the Human-Machine Handoff

In projects like this, the hardest part is often not invoking the model, but transferring responsibility between the model and human reviewers.

The first error class is incorrect notation transcription. A changed subscript, quantifier, or Greek letter can leave the surface structure fluent while altering the mathematical meaning.

The second is missing context. A definition in Gödel’s manuscripts may appear dozens of pages earlier or rely on notation from an earlier paper. If the model works from an isolated passage, it may reconstruct the relationship between references incorrectly.

The third is a broken proof chain. A model may treat “previously established” as “derivable here,” or mistake a planned statement in a draft for a completed argument.

AI output therefore cannot be treated as proof that a theorem holds. A sound workflow needs at least three layers of review:

  1. Compare the transcription with the original scans item by item;
  2. Use proof assistants such as Lean or Coq to check the machine-verifiable portions;
  3. Have domain experts assess unformalized steps and verify bibliographic sources.

Formal verification is not infallible either. Every definition entered into a proof assistant must be correct, and each imported lemma must fit the original context. A proof assistant can validate an encoded proof, but it cannot automatically determine whether the encoded object was understood correctly.

In other words, “the machine verified it” and “a mathematician confirmed its fidelity to the manuscript” are separate checks.

From Scanned Attachments to Searchable Knowledge

The output should not be described merely as a “discovery.” Its more valuable deliverable is a searchable, traceable, and reusable manuscript database.

In the past, a historian of mathematics might never encounter a repeated argument written in the margin of a page. After transcription and cataloging, it has at least a chance of entering a full-text search system. Researchers can search for a definition, a proof strategy, or a body of arguments that recurred without ever being published.

That creates several direct benefits:

  • Experts no longer need to repeatedly purchase, scan, and manually transcribe the same archives;
  • Unpublished results can be annotated and linked to later papers;
  • History of mathematics can draw on finer-grained chains of evidence than paper-based catalogs provide;
  • Teaching materials can use authentic drafts to show how proofs develop, rather than presenting only polished final versions.

The same dataset may also support training, retrieval-augmented generation systems, or automated benchmarks. That stage requires additional work on provenance, permissions, attribution, and normalization of historical mathematical terminology. Manuscript scans should not simply be dropped into a training pipeline and the resulting output labeled a new discovery.

The database is the infrastructure. The model is one of the tools used to build and work with it.

Source

The Guardian: AI helps unlock hidden discoveries in the man who transformed maths

Editorial Assessment

This development is better understood as an upgrade to research infrastructure than as an independent AI breakthrough in mathematics. Mathematicians once spent substantial time identifying scans, assigning numbers, and locating material by hand. Models can now begin with these tedious lower-level tasks, while problem selection, contextual judgment, and proof-chain verification remain the scarce skills.

The first clear gains are most likely to appear in transcription, cataloging, and formalization rather than in fully unsupervised original proofs. The former can be checked item by item against the source, making errors easier to locate. The latter requires outside research judgment even before deciding what is worth proving. In the coming months, the important measures will be whether the manuscript database becomes publicly accessible, how much of it receives expert review, and which results enter peer review or formal proof libraries.

This differs from earlier “AI solves math” stories because the material comes from historical manuscripts and every stage of the workflow can be compared with the scans; there is no need to accept the model’s conclusion on faith. The original report, however, described the collection as reportedly containing hundreds of proven theorems. Without checking them individually, it is not possible to determine the formalization coverage from the report alone. The project shows that AI can lower the barrier to using old manuscripts, but it does not yet show that AI has changed how mathematical discoveries are divided between people and machines.

Source: https://www.theguardian.com/science/2026/nov/13/kurt-godel-notebooks-ai-helps-unlock-hidden-discoveries-man-who-transformed-maths