Self-supervised models learn robust representation of real images under cropping and resizing, which can be applied to detect AI-generated images without training.