The article discusses the emergence of generative AI-generated menus in the restaurant industry, which often produce illustrations that appear eerily flawless, symmetrical, and smooth, evoking a sense of unease in observers. These AI-generated images can range from obviously fake, such as a burrito with exaggeratedly melty cheese, to subtly off-putting, where the irregularities are only noticeable upon closer inspection. The underlying cause is attributed to the training data used for AI models like large language models (LLMs) and diffusion models, which are trained on vast datasets that include a narrow, “pleasing” aesthetic, leading to a homogenized output.
The article explains that AI models are trained on extensive datasets, identifying patterns to predict user requests, such as generating a menu for a burger restaurant. However, when these models are trained on too much of their own AI-generated content, they risk “model collapse” or “convergence,” where the quality of outputs degrades without becoming entirely useless. This results in AI-generated menus that mimic the styles of popular fast-food chains, further reinforcing a uniform aesthetic. Additionally, the optimization of datasets for “pleasingness” contributes to a smoothing effect, making images increasingly idealized and less realistic with repeated edits. This phenomenon aligns with the “uncanny valley” effect, where near-realistic AI-generated food images elicit disgust and unease, intensifying public backlash against AI-generated menus in restaurants. The implications extend beyond the culinary realm, raising concerns about trust in visual evidence in broader societal contexts. Read the full article here.