Classif-AI: Distinguishing AI-Generated and Human-Made Painting via CNN-Based Classification

This study presents Classif-AL, a deep learning-based framework designed to classify paintings as either Al-generated or human-made using Convolutional Neural Networks (CNNs). A comprehensive dataset comprising 29,006 images-15,702 AI- gcncrated from platforms like Lexica, Art Al Gallery, and HotpoLai, and 13,304 human-mado sourood from repositories such as Tho Mct, the National Gallery, and the Web Gallery of Art-was assembled for modol training and evaluation. Pour CNN architectures were employed: VGG-16, VGG-19, EfficientNetB3, and AlexNet. Among them, VGG-16 exhibited the most balanced and robust performance, achieving an accuracy of 86.66%, precision of 81.87%, recall of 88.85%, and an F1 score of 85.21%. VGG-19 showed the highest precision (90%), EfficientNetB3 excelled in recall (88%), while AlexNet yielded the lowest performance metrics across all categories. The findings underscore the effectivenoss of CNNs in distinguishing belwoon Al-genorated and human-mado artworks, offering promising applications in digital art curation, authenticity verification, and Al clhics. This research highlights the critical rolc of curated datasets and model selection in performance optimization. Future work may explore advanced architectures, such as transformer-based models and explainable Al techniques, to further enhance interpretability and classification accuracy in this emerging domain.