The democratisation of AI image generation has enabled people to create images with just a simple text prompt. Although beneficial in particular areas such as creative processes or business marketing, it has also given way to more harmful and dangerous acts such as identity theft, the viral spread of misinformation and disinformation, and the non-consensual creation of deepfake content. The expansion from GAN-based to diffusion-based image generative models adds an additional dimension to the problem. As humans find it more difficult to detect these images, so do detection models designed to identify synthetic imagery in the frequency domain. Detectors optimised for GAN-specific frequency artifacts have shown limited ability to generalise to the fundamentally different spectral characteristics produced by diffusion models, creating a gap between their detection capabilities and the diversity of existing generative architectures. This project investigates the generalisation of a GAN-optimised image detector to unseen and modern generative architectures, and the extent to which GAN-trained frequency-domain detectors fail to generalise to unseen generators, including but not limited to diffusion-based models.