The Uniformity Issue Behind Unappealing AI-Generated Menus

Lisle notes that the architecture of these models can help clarify why the illustrations tend to adopt such a particular aesthetic — one where each ice cream scoop is perfectly spherical, and shrimp seem to have been genetically altered to consume their own tails, giving rise to new “Lovecraftian food horrors.”

Large language models (LLMs) and diffusion models — the types of AI models that power seemingly all-knowing chatbots and image generators like ChatGPT and Midjourney — are trained on extensive datasets. The models then discern patterns within this data to anticipate what a user is requesting when they query something like, “Create a menu for a burger restaurant.”

“A lot of this looks like a Chili’s menu from 2015, and there’s a reason for that,” Lisle stated. “That was the source material from which [the models] derived their functionality.”

For companies developing AI models, new training data is crucial — Amazon has even been found to seek out rare books for scanning and adding to its training data, subsequently destroying the physical copies after uploading them. It’s inevitable that some AI-generated material will infiltrate these incomprehensibly vast datasets. However, when AI models begin to rely too heavily on their own AI-generated content, they risk encountering model collapse.

“Model collapse is somewhat akin to a mad cow disease… when you feed a model’s outputs back into itself, eventually the inbreeding becomes too much, and the entire system collapses,” Lisle elaborated. “What we observe here is convergence, which isn’t necessarily model collapse.”

Convergence is somewhat less severe, compromising the quality of an AI’s outputs without rendering them completely ineffective.

If a user prompts an AI model to generate a menu for a fast food restaurant, it will likely draw from the menus of Wendy’s, Burger King, McDonald’s, or another well-known chain. These menus typically share a similar design, meaning that the AI-generated results will replicate that style, further reinforcing it if the AI-created menu is later included in the training data.

Yet menus and food advertisements will consistently appear more attractive than the actual items, similar to a Big Mac in a McDonald’s ad, where each ingredient of the sandwich is arranged by a prop stylist to maximize visual appeal. This phenomenon can be even more exaggerated in AI results.

“The optimization of the datasets leans towards pleasing aesthetics, or, you know, avoiding any offensiveness, which leads to a kind of homogenization,” Lee Rainie, Director of the Imagining the Digital Future Center at Elon University, shared with TechCrunch. “AI tends to smooth out the rough edges, both in images and language.”

On a more localized level, this smoothing effect appears when using an AI image generator to create a menu and then modify it. On X, a user named Labtec demonstrated what occurs when you generate a menu in ChatGPT and then edit it 100 times, causing the food to progressively look less authentic. (We replicated the experiment and observed similar outcomes.)

“The final product actually gives me an unsettling feeling,” Labtec commented.

Restaurants may be succumbing to this issue, continuously modifying their AI-generated menus to tweak minor details, such as prices or item titles. It appears that with each modification, the food images become slightly rounder and smoother.

“People have an almost indefinable instinct about when they’re looking at something AI-generated versus something that originated from reality,” Rainie noted. “There’s a sensitivity that some individuals find challenging to articulate, yet they recognize it when they see it. I believe this contributes to the early backlash stories regarding restaurants utilizing AI menus.”

There is scientific backing for our aversion to these AI menus. Researchers at the University of Duisburg-Essen in Germany discovered that AI-generated food images produce an “uncanny valley” effect, where images of food appearing nearly real provoke more disgust and discomfort than those that are obviously artificial. This discomfort heightens in light of the current cultural context surrounding AI.

If people have such strongly negative reactions to these images, it stands to reason that restaurants might reconsider their attempts to implement AI menus. However, the complications that create flawlessly browned hamburger buns extend beyond just dining.

“Seeing and hearing have always been synonymous with belief, to the extent that even our legal systems rely heavily on the notion that the gold standard for evidence is recorded confessions and videotaped proofs,” Lisle remarked. “That’s no longer applicable. The world has undergone a fundamental change, for better or worse.”