As AI usage in mathematics grows, distinguishing genuine AI involvement in proof construction from mere AI assistance in writing becomes crucial. We leveraged historical arXiv papers, recent works that openly disclose substantive AI use, and manuscripts released by OpenAI to build five cohorts comprising 8,281 papers, including AI‑generated drafts and matched pairs addressing the same research problem. We then introduced the ProofTrace framework, which represents each paper as a Graph of Immediate Proof Support (GIPS) and extracts interpretable features of mathematical organization. Our findings are threefold: 1) Principal‑component projections of the features form a qualitative gradient from human‑written, through AI‑reworded, to AI‑generated and finally to papers that disclose substantive AI use; 2) AI‑generated proofs exhibit recurring signatures such as fewer external citations, heavier reliance on intermediate conclusions, and altered graph connectivity—signatures that remain significant after controlling for research field, manuscript length, or graph size; 3) In a separate detection experiment, ProofTrace reliably separates purely human papers from those disclosing substantive AI use, maintaining performance under AI re‑wording where state‑of‑the‑art text‑based detectors degrade sharply. $$\text{GIPS} = (V, E),\; V\text{ denotes proposition nodes},\; E\text{ denotes direct support edges}$$ Review: The study offers quantifiable metrics of AI involvement, providing a practical tool for peer review and academic integrity monitoring.