At Chi-Town Deli in Chicago, artificial intelligence was given a mission that seemed almost impossible to get wrong: make some sandwiches look appetising. The result demonstrated once again that technological progress does not always move in the direction of the human stomach. Chris Murphy, who had stopped by to buy lunch, immediately noticed that the menu photographs had been generated with AI. Every sandwich had the same artificial shine, the lettuce looked as though it had been built from bricks, and one Reuben had acquired an appearance Murphy described as “Lovecraftian.” Other menus that went viral have featured suspiciously perfect round shrimp, burritos with almost organic textures and ingredients that appear to have continued evolving after the cook went home. The restaurant manager, Ali Malik, explained that the images had been made by an employee and were only temporary, pending the arrival of a display featuring real photographs. You save money on the photographer, the lighting, the food styling and the camera. The only problem is that you must then convince the customer that the real sandwich will not look like the picture.

Why does AI make precisely these kinds of mistakes? Most modern image generators are diffusion models. They begin with visual noise and gradually reconstruct an image that is statistically compatible with the prompt and with the images seen during training. The model does not, however, possess an “internal recipe” for a sandwich, nor does it check whether every lettuce leaf, slice of meat or shrimp forms a physically coherent object. This is how structural consistency errors appear: elements merge together, ingredients repeat unnaturally, shapes acquire impossible geometry and textures continue from one object into another. There is also texture repetition, the model’s tendency to reproduce local visual patterns. This can make lettuce resemble bricks, holes or scales. Errors involving specular highlights can also produce exaggerated reflections and shine, giving food that wet, plastic and artificial appearance that is so easy to recognise. New models have become very good at the overall composition of an image, but they can remain surprisingly weak at local coherence. From a distance, they see a sandwich. Look more closely and you may discover that some of its ingredients appear to belong to a species yet to be described by biology.