Food images in restaurant menus created by AI are drawing attention with their flawless symmetry, excessive smoothness, and unsettling quality despite looking realistic. Reality Defender CTO Alex Lisle compares these images to an alien trying to make pizza without understanding its basic principles. In some examples, cheese that is excessively puffed up or shrimp that appear to be eating their own tails makes their artificial origin immediately apparent, while in most images, the problem becomes noticeable only upon closer inspection.
According to Lisle, this aesthetic is linked to how AI models are trained. The large language models and diffusion models used in systems such as ChatGPT and Midjourney learn patterns from enormous datasets and try to generate content that matches users’ requests. Therefore, when asked for a burger restaurant menu, they may draw on the similar menu styles of chains such as Wendy’s, Burger King, or McDonald’s. Lisle also attributes the reason why the resulting images sometimes look “like a Chili’s menu from 2015” to the dataset that feeds the models.
AI-generated content being reintroduced into training data can cause models to feed on their own outputs. When this situation becomes excessive, it creates the risk of “model collapse.” Lisle compares this to the genetic inbreeding that occurs when a model’s outputs are repeatedly fed back into the model itself. However, he says that what is seen in menus is not direct model collapse, but rather “convergence,” which reduces output quality. Lee Rainie, director of the Imagining the Digital Future Center at Elon University, says that optimizing models to favor non-offensive and “likable” content leads to homogenization by smoothing out distinctive features in images and language.
In an experiment conducted by an X user named Labtec, a menu created with ChatGPT was edited 100 times, and the food images were found to become increasingly rounder and smoother. TechCrunch repeated a similar experiment. It is noted that restaurants repeatedly editing menus to change prices or product names could produce the same effect.
Researchers at the University of Duisburg-Essen in Germany found a “uncanny valley” effect in AI-generated food images. Images that look almost real create more disgust and unease than those that are clearly fake. This reaction combines with a broader cultural distrust of AI-generated content. Lisle says the era in which images and audio recordings were accepted as evidence without question is over, and that this affects broader areas, including courts.
Why it matters
These findings show that when preparing menus, restaurants need to consider not only visual quality but also the uniformity that repeated processing of AI outputs may create. As product names and prices are changed, images becoming increasingly similar could weaken the elements that distinguish businesses’ brands and affect customers’ trust in menu photos. The fact that the problem is not directly model collapse but convergence caused by models gravitating toward similar and appealing examples suggests that the solution may not be limited to using more data. The unease created by food that looks almost real shifts the debate away from aesthetics and toward the question of how AI-generated images should be evaluated in advertising, as evidence, and in everyday communication.