Restaurant menus featuring AI-generated food illustrations are being criticised for their uniform and unappetizing appearance. Customers reportedly experience a 'visceral sensation' that something is amiss with these images, which often look 'eerily flawless' and 'precisely symmetrical'.
Alex Lisle, CTO of Reality Defender, suggests that the specific aesthetic stems from how these AI models are built. He notes that much of the training data for these models, including large language models (LLMs) and diffusion models, references existing menus from popular fast-food chains, leading to a homogenised style.
Lee Rainie, Director of the Imagining the Digital Future Center at Elon University, explained that the optimisation of datasets for 'pleasingness' or 'not being offensive' results in a 'homogenization' where AI 'shaves off the edges' of images. This effect can intensify with repeated edits to AI-generated images, making food look increasingly round and smooth.
Research from the University of Duisburg-Essen in Germany indicates that AI-generated food images can elicit an 'uncanny valley' effect, where images that are almost real cause more disgust and unease than those clearly artificial. This aversion contributes to a backlash against restaurants using AI menus.