Art has long been recognized as a central medium of cultural expression, education, and human creativity. However, the rapid enhancements of advanced AI tools such as Stable Diffusion, Midjourney, and DALL·E has introduced large volumes of synthetic images that disguise the boundaries between authentic human artwork and machine-generated creations. This need raises critical challenges for evaluating the clarity and authenticity of digital art, particularly in distinguishing whether a piece is real or artificially generated. Traditional image analysis methods, primarily relying on traditional features or convolutional neural networks (CNN), limitations in capturing both fine-grained textures and global artistic composition. To address this challenge, this study proposes a Memory-Augmented Vision Transformer (MeMViT), an integrated EfficientNet-based mechanism works using hierarchical transformer model enhanced with masked attention and memory tokens to efficiently capture both local and global image dependencies. The model was trained on a balanced dataset of 60,000 images, comprising equal proportions of AI-generated and real artwork. For comparison, several baselines, including CNN, DenseNet, AlexNet, and a standard ViT, were also evaluated. Experimental results demonstrate that the proposed model achieves the highest classification accuracy of 97%, outperforming all baselines. Furthermore, model interpretability was ensured using Grad-CAM visualizations and LIME analysis, providing transparent insights into decision-making for art clarity evaluation which shows the AI’s impact on creative authenticity.
Jiexiao et al. (2026) studied this question.