Study Shows AI Develops Distinct Proof Style When Reproving Euclid’s Elements

The Grafy Research Lab released a paper on November 13, 2026 that asks whether large language models (LLMs) develop a proving style that is recognizably different from human mathematicians. To answer this, the authors asked three LLMs to independently reprove every one of the 465 propositions in Euclid’s Elements and then compared the AI‑generated citations to the human‑proof baseline using Jaccard similarity and complex‑network analysis (according to the arXiv preprint).

What was the goal of the research?

The study set out to determine if AI‑generated mathematical arguments form a coherent stylistic signature, separate from the conventions that have evolved over two millennia of human mathematics. As more researchers rely on LLMs to draft or check proofs, understanding any systematic differences becomes crucial for interpretability, trust, and collaboration between humans and machines.

How did the authors generate and evaluate the AI proofs?

Three state‑of‑the‑art LLMs were each tasked with producing a full proof for every proposition in Euclid’s Elements. The models were run without any human‑provided hints, ensuring that each proof was created solely from the model’s internal knowledge. After generation, the authors extracted every citation the AI used—whether referencing earlier propositions, external theorems, or textbook sections. These citation sets were then compared to the citation choices made in the canonical human proofs. Two quantitative tools were employed:

1. Jaccard similarity – measuring the overlap between AI and human citation sets for each proposition.
2. Complex‑network analysis – treating citations as edges in a graph to see how AI‑generated networks differ in structure from the human baseline.

What did the analysis reveal about AI‑generated proof style?

The authors report that AI‑generated proofs exhibit citation patterns that diverge from the human baseline, indicating a distinct proving style. While the paper does not quantify the exact degree of divergence in the abstract, the use of Jaccard similarity and network metrics confirms that the differences are systematic rather than random. In other words, the models are not simply copying human reasoning; they are constructing proofs with their own internal logic for selecting supporting statements.

Why does this matter for creators, developers, and researchers?

For mathematicians

The finding suggests that AI‑assisted proof drafting may introduce novel argument pathways that differ from traditional textbook exposition. This can be both an opportunity—offering fresh insights—and a risk, because unfamiliar citation choices can make AI‑generated proofs harder to verify and trust.

For AI developers

Identifying a unique citation signature provides a concrete target for model improvement. Developers can use the metrics from this study to fine‑tune models toward more human‑like citation behaviour, or deliberately preserve the AI’s distinct style if it proves beneficial for discovery.

For educators and policy makers

Understanding that AI can produce mathematically correct but stylistically atypical proofs helps shape guidelines for the responsible use of generative models in academic settings. Transparency about how a proof was generated becomes essential.

How can Grafy users explore or apply these insights?

Grafy offers tools that let you replicate parts of this workflow without writing any code:

• Reader (/reader) lets you upload the arXiv PDF, ask natural‑language questions about the methodology, and retrieve specific sections such as the Jaccard similarity results.
• Chat (/chat) lets you run multiple LLMs side‑by‑side on the same mathematical prompt, observe differences in the generated proofs, and even request the models to list the citations they would include.
• Graph (/graph) can capture the branching prompts and outputs, allowing you to visualise how different models evolve their citation networks over successive propositions.

By combining these apps, a researcher can reproduce a small‑scale version of the experiment—say, reproving the first ten propositions—and immediately see the citation graphs that emerge. This hands‑on approach turns the paper’s findings into an interactive learning experience.

What are the next steps for the community?

The study opens several avenues for further work:

1. Scaling to other mathematical corpora – applying the same analysis to modern textbooks or research papers to see if the AI style persists.
2. Human‑AI collaboration studies – measuring how a mixed workflow (human writes part of the proof, AI fills gaps) affects citation similarity.
3. Model‑level diagnostics – using the citation‑network signatures as a diagnostic tool to gauge a model’s readiness for formal mathematics tasks.

Grafy’s modular environment makes it straightforward to set up these experiments, especially with the Operator browsing agent for pulling in external datasets and the Code studio for running statistical analyses on citation graphs.

Frequently asked questions

Does the paper claim AI proofs are more accurate than human proofs?

No. The research focuses on stylistic differences in citation choices, not on overall correctness or superiority of the AI‑generated proofs.

Can I use Grafy to automatically compare citation similarity for my own proofs?

Yes. By uploading your proof set to Reader and prompting a model in Chat to generate comparable proofs, you can then extract citation lists and compute Jaccard similarity manually or with a small script in Code.

Are the three LLMs used in the study publicly available?

The paper does not name the specific models, only that they are “large language models.” Grafy’s Chat app provides access to many contemporary models, allowing you to experiment with similar capabilities.

---

Source: Grafy Research Lab — AI has its own style in mathematical proving