\n<\/p>\n
When it first happens to you, you think you\u2019re crazy. You wander into a cafe and look at a menu with a variety of bagel sandwiches, but each illustration looks eerily flawless, precisely symmetrical, and oddly smooth, eliciting a visceral sensation that something isn\u2019t right. You might think you\u2019re paranoid, but you\u2019re not losing your mind. Generative AI menus have hit the restaurant business courtesy of models courtesy of models trained on a narrow, \u201cpleasing\u201d aesthetic that produces a look that feels wrong even when you can\u2019t articulate why.<\/p>\n
Sometimes, these illustrations are egregiously fake, like a burrito with cheese so bubbly and melty that it looks more like avant garde art<\/a> than lunch. More often, they\u2019re so ordinary<\/a> looking<\/a> that you only notice something is wrong<\/a> when you take a second to look more closely.<\/p>\n \u201cIt\u2019s almost like an alien trying to make a pizza without understanding its core principles,\u201d Reality Defender CTO Alex Lisle told TechCrunch. (Reality Defender itself is part of a growing category of startups selling AI-detection and content-verification tools \u2014 a business that exists in part because of issues like this one.)<\/p>\n Lisle says that the way these models are built can help explain why illustrations seem to embrace such a specific aesthetic \u2014 one where every ice cream scoop is perfectly round, and where shrimp seem to have been genetically modified to eat their own tails, creating new \u201cLovecraftian food horrors<\/a>.\u201d<\/p>\n Large language models (LLMs) and diffusion models \u2014 the kinds of AI models that make seemingly omniscient chatbots and image generators like ChatGPT and Midjourney possible \u2014 are trained on vast quantities of data. The models then identify patterns in the datasets to predict what a user is looking for when they ask something like, \u201cMake me a menu for a burger restaurant.\u201d<\/p>\n \u201cA lot of this stuff looks like a Chili\u2019s menu from 2015, and there\u2019s a reason for that,\u201d Lisle said. \u201cThat was the corpus of work from which [the models] drew their function.\u201d<\/p>\n New training data is invaluable to the companies building AI models \u2014 Amazon has even been found to source rare books to scan and add to its training data, only to destroy those books<\/a> once they\u2019ve been uploaded. It\u2019s inevitable that some AI-generated content will seep into these incomprehensibly large data sets. But when AI models train on too much of their own AI-generated content, they risk model collapse<\/a>. <\/p>\n \u201cModel collapse is almost like a mad cow disease\u2026 when you feed the outputs from one model back into itself, eventually the inbreeding becomes too much, and the whole thing collapses,\u201d Lisle explained. \u201cWhat we see here is convergence, which isn\u2019t necessarily model collapse.\u201d<\/p>\n Convergence is a bit less extreme, degrading the quality of an AI\u2019s outputs without making it entirely useless. <\/p>\n If someone asks an AI model to generate a menu for a fast food restaurant, the model will likely reference menus from Wendy\u2019s, Burger King, McDonald\u2019s, or another popular chain. These menus already share a similar style, which means that the AI-generated outputs will mimic that same style, only to further reinforce it further if the AI-generated menu ends up back in training data.<\/p>\n But menus and advertisements for food will always look better than the real thing, like a Big Mac in a McDonald\u2019s commercial where each layer of the sandwich is arranged by a prop designer to look maximally appetizing. This effect can become even more pronounced in AI outputs. <\/p>\n \u201cThe optimization of the data sets is for pleasingness, or you know, not being offensive, and so there\u2019s a way that turns into homogenization,\u201d Lee Rainie, Director of the Imagining the Digital Future Center at Elon University, told TechCrunch. \u201cWhat AI is known to do both in images and language is to shave off the edges.\u201d<\/p>\n On a more localized scale, this smoothing of images seems to happen when you use an AI image generator to create a menu and apply edits to it. On X, a user named Labtec<\/a> showed what happens when you make a menu in ChatGPT, then edit it 100 times to see how the food continues to look less and less like it should. (We replicated the experiment and found similar results.)<\/p>\n \u201cThe end result actually makes me uncomfortable,\u201d Labtec wrote.<\/p>\n