Detecting AI Signatures in Mathematical Proofs: What the AI4Math Study of 8,281 Papers Found

Quick answer

• A Harvard research team has published a method for detecting signatures of AI involvement in mathematical proofs. The paper "The Gold Rush in AI4Math, II: What are signatures of AI-assisted Proofs?" by Yuanchuan Guo, Jiashun Jin, Zheng Tracy Ke and Bingcheng Sui was submitted to arXiv on 5 November 2026 (arXiv:2611.08412).
• The study analyses 8,281 papers across five cohorts with different levels of AI involvement. The cohorts include historical human-written papers, AI-reworded papers, papers disclosing substantive AI use, AI-generated (rebuilt) papers and manuscripts released by OpenAI, including matched pairs of papers addressing the same research problems.
• The AI signatures live in the logical structure of a proof, not in its writing style. AI-generated proofs tend to cite fewer externally sourced results, rely more heavily on intermediate conclusions they create themselves, and show altered graph connectivity compared with human proofs.
• The ProofTrace framework keeps its detection performance after AI rewording. In the authors' detection experiment, ProofTrace remained stable under AI rewording, whereas state-of-the-art text-based detectors declined sharply.
• This is a new arXiv preprint. Every result reported here was published by the authors in the paper itself and has not yet been independently peer-reviewed.

What are signatures of AI-assisted proofs?

Signatures of AI-assisted proofs are measurable characteristics of how a proof is organised — how its theorems, lemmas and intermediate conclusions link to and depend on one another — from which the level of AI involvement in producing that proof can be inferred, separately from AI assistance with wording alone.

Unlike conventional AI-text detectors, which rely on word choice and sentence rhythm, the approach in this study represents each paper as a graph of immediate proof support (GIPS): every stated result is a node, and an edge indicates that one result is used directly to derive another. The framework, called ProofTrace, extracts interpretable features of mathematical organisation from this graph.

Who carried out this AI4Math study, and what data did they analyse?

The paper was written by Yuanchuan Guo, Jiashun Jin, Zheng Tracy Ke and Bingcheng Sui, as part of "The Gold Rush in AI4Math" series, and was submitted to arXiv on 5 November 2026 in the Artificial Intelligence category (cs.AI). According to the Facebook post by the Page "Học lập trình cho người mới bắt đầu" that shared the study, the team is based at Harvard.

First page of the paper "The Gold Rush in AI4Math, II: What are signatures of AI-assisted Proofs?" on arXiv (arXiv:2611.08412), showing the title, author list and abstract. Image from the original Facebook post.

The authors constructed five cohorts comprising 8,281 papers in total, combining existing manuscripts with manuscripts they produced themselves using generative AI. The source material includes historical arXiv papers, recent papers disclosing substantive AI use, and manuscripts released by OpenAI. The cohorts span different levels of AI involvement and include matched pairs of papers addressing the same research problems — a case-control element that earlier work lacked.

Which three structural features reveal AI signatures in a proof?

Across 700 matched pairs — in each pair, an AI-rebuilt paper tackles the same research problem as its human-written counterpart — the authors measured three structural features and found consistent differences:

Structural feature	Human-written proofs	AI-rebuilt proofs	What it means
Fraction of results cited from external sources	6.1%	0.2%	AI proofs rely far less on pre-existing results outside the paper
Fraction of stated results that directly use an intermediate step	58%	87%	AI proofs lean much more on intermediate conclusions they generate themselves
Connected components per node (separate proof strands, ignoring arrow direction; 699 pairs)	0.24	0.14	The connectivity structure of AI proof graphs differs measurably from human ones

These figures are reported by the authors in the paper (summary over the 700 matched pairs; the third feature is computed on 699 pairs). According to the authors, many of these signatures persist under controls for research field, manuscript length or graph size.

When the features are projected onto principal components, a qualitative gradient of AI involvement appears: AI-reworded papers remain close to human papers, the two AI-generated cohorts depart in broadly shared directions, and papers disclosing substantive AI use occupy intermediate positions that shift toward the AI-generated cohorts over time.

Why do text-based AI detectors fail after rewording?

Text-based detectors measure signals on the surface of language: vocabulary, sentence rhythm, unusual fluency. When a paper is reworded by AI, that surface changes almost completely, and the detector's signal disappears with it.

ProofTrace does not measure the language surface; it measures the organisation of the proof — which results cite outside sources, and which results depend directly on which intermediate conclusions. Rewriting the wording does not change those dependency relations, because they belong to how the argument is built, not to how it is phrased. This is why, in the authors' separate detection experiment, ProofTrace distinguished human-written papers from those disclosing substantive AI use and kept its performance stable under AI rewording, while state-of-the-art text-based detectors declined sharply.

What does this study mean for mathematics and AI research?

First, the study separates two questions that are often conflated: whether AI helped rewrite a text, and whether AI participated in the proof itself. The second question is the one that matters for research integrity, and it can only be answered by looking at the structure of the proof.

Second, the method offers journals, conferences and research communities an interpretable direction for verification: instead of an unexplained "AI-likeness" score, features such as the external-citation fraction or the reliance on intermediate conclusions can be inspected and debated concretely.

A balanced reading is still required: this is a new preprint posted on arXiv in November 2026, and the results are the authors' own published findings. The method's lasting value will depend on independent replication and critique — the way science treats every new result.

Frequently asked questions about AI signatures in mathematical proofs

How many papers did the AI-signatures study analyse?

The study analysed 8,281 papers in five cohorts spanning different levels of AI involvement, including 700 matched pairs in which an AI-rebuilt paper addresses the same research problem as a human-written paper.

What is ProofTrace?

ProofTrace is the framework proposed by the authors in arXiv:2611.08412. It represents each paper as a graph of immediate proof support (GIPS) — stated results are nodes, and a "used directly to derive" relation is an edge — then extracts interpretable features of how the proof is organised.

Can AI rewording a proof in a human-like style fool the detection?

According to the detection experiment reported in the paper, AI rewording sharply reduces the performance of text-based detectors, but ProofTrace's performance remains stable, because ProofTrace measures the dependency structure of the proof rather than its wording.

What are the three typical structural signatures of AI-generated proofs?

According to the authors, the three signatures are: fewer results cited from external sources, a higher fraction of results relying directly on intermediate conclusions, and altered graph connectivity (fewer separate proof strands per node in the measured matched pairs).

Has this research been independently verified?

Not yet. The paper is a preprint posted on arXiv on 5 November 2026; the results reported here are the findings published by Guo, Jin, Ke and Sui in the paper itself, and independent replication and peer review will take time.

Sources

This article was rewritten from a post by the Facebook Page Học lập trình cho người mới bắt đầu (published 9 November 2026): original Facebook post. Original scholarly source: Yuanchuan Guo, Jiashun Jin, Zheng Tracy Ke, Bingcheng Sui, "The Gold Rush in AI4Math, II: What are signatures of AI-assisted Proofs?", arXiv:2611.08412, submitted 5 November 2026 — read the paper on arXiv. All figures in this article (8,281 papers, 700 matched pairs, the ratios 6.1%/0.2%, 58%/87%, 0.24/0.14) are taken from the abstract and results summary of the paper itself on arXiv.