Multi-metric plagiarism scale for AI paintings

Generative Adversarial Networks (GANs) can now generate images with human-like textures, styles, and compositional structure, complicating originality verification and blurring the line between legitimate influence and impermissible copying. Existing plagiarism and similarity methods—largely designed for text or conventional image workflows—often fail to reflect the aesthetic, statistical, and generative characteristics of GAN outputs, creating evidentiary gaps in attribution and accountability. This study targets AI-replicated realist (representational) paintings and reports similarity as a graded plagiarism scale (CSI), not a binary verdict. This study proposes an experimentally validated, transparent multi-metric framework for assessing potential plagiarism in GAN-based art. Five complementary indicators are integrated into a composite similarity score: Structural Similarity Index (SSIM) for global structure, Gram-matrix statistics for texture, Earth Mover’s Distance (EMD) for distributional shifts, Canny edge maps for contour correspondence, and HSV histogram comparisons for color affinity. Using a curated dataset that pairs original artworks with GAN-generated derivatives—ranging from near-reproductions to subtle stylistic overlap—we evaluate each metric’s individual performance and their combined effectiveness.

Results show that no single metric reliably detects plagiarism across diverse stylistic conditions. In contrast, the composite score improves detection accuracy, robustness to style variation, and interpretability, supporting more consistent screening and forensic review. We further report normalization procedures, cross-validated weighting, threshold calibration for decision support, and uncertainty estimates to enhance auditability and reproducibility. The proposed scale offers a methodologically sound basis for flagging suspect works and for reporting similarity evidence in artist, research, and legal contexts.