The Silent Brush: Artistic Style Leakage in AI Art Generation

Generative text-to-image models often produce outputs that appear novel yet reflect stylistic patterns learned from training data. While prior work suggests these models internalize style as statistical regularities, systematic methods to measure such influence remain limited. We introduce Art Arena, a protocol for quantifying stylistic influence through three stages: Entry Trials test for stored stylistic traces under explicit attribution, Motif Duels probe interactions and hybridization under controlled prompts, and the Influence Ledger ranks styles by their likelihood to reappear when style is not mentioned—a phenomenon we term The Silent Brush. By converting influence into an empirical signal, Art Arena enables auditing of stylistic leakage and provides insight into how training data shapes generative behavior. While developed for text-to-image systems, the approach generalizes to other modalities and encourage transparency in creative AI.