Restaurants are increasingly turning to AI-generated illustrations for their menus, but the results often look subtly "off" — unnaturally smooth, symmetrical and artificial in ways diners struggle to articulate but instinctively notice. Experts say this is because image-generating AI models are trained on huge datasets of existing food imagery, which pushes outputs towards a narrow, formulaic "pleasing" aesthetic rather than anything resembling real food, producing oddities such as perfectly round ice cream scoops or shrimp that appear to eat their own tails.

Alex Lisle, chief technology officer at AI-detection firm Reality Defender, explained that many of these images echo dated commercial styles, such as "a Chili's menu from 2015," because that is the kind of imagery the models were originally trained on. He also warned of a related phenomenon called "convergence," distinct from full "model collapse": when AI-generated menus for chains such as Wendy's or McDonald's mimic existing similar-looking marketing material, and those AI images then feed back into future training data, the sameness becomes self-reinforcing, further homogenising how AI depicts food over time.