Toss a rock anywhere in New York City these days, and you may well strike an AI-generated food ad. That, at least, was the verdict of X user Maggie Moda, who documented a series of odd sidewalk signs outside the city's delis and cafés in a post that has since drawn 2.8 million views. The images included croissant sandwiches with almost comically perfect layers, an affogato dotted with random holes, and a chopped-cheese hero with cheese melted a bit too flawlessly over every piece of meat.

Generative AI now offers restaurants a cheap and remarkably easy alternative to photographing the food they actually sell. For small businesses that lack the time or budget for professional food photography, that can be a genuine advantage. Yet the backlash to AI food ads exposes a paradox: some customers would rather see an imperfect picture of a meal that really existed than a polished image of one that never did.

That distinction matters beyond the question of whether AI food simply looks unappetizing. Restaurants have been making their food look better in advertisements for decades, but generative AI makes it possible to advertise a photorealistic meal that never existed at all. Because that imaginary meal is being used to sell a real product to real customers, it raises an old question in a new form: how different can the food in the ad be from the food on the plate before the ad becomes deceptive? It is a question consumer-protection regulators have long wrestled with—advertising in the United States is policed by the Federal Trade Commission under its truth-in-advertising rules—and it is now being asked about images produced by tools that are increasingly woven into how small businesses market themselves.

The closer you look at an AI food ad, the less it resembles lunch—and the more it starts to trigger trypophobia, the fear of patterns of holes. Sandwiches contain identical holes and endlessly repeating ingredients. One viral burrito featured so many tiny holes that the internet christened it the "trypophobia burrito."

In other ads, cheese melts without any obvious beginning or end, ingredients multiply into unnaturally neat rows, and croissants achieve levels of flakiness that appear to defy structural engineering. The images have become recognizable enough to earn their own unappetizing nickname: food slop—and not the sloppy Joe kind, but AI slop.

The Business Case for AI Slop

For restaurants, though, the appeal is not hard to grasp. Professional-looking food advertising no longer necessarily requires professional food photography.

When Grind & Unwind opened on San Francisco's Haight Street in May, owners AJ and Lyndsey Lozano placed a temporary menu in the window featuring AI-generated images of the restaurant's food while they got the new business up and running, according to SFGATE. The internet was not impressed. A Reddit post titled "Yum, slop" drew more than 670 upvotes, with commenters comparing the focaccia to a "loofah" and a "lung."

The reason behind the AI images, however, was far less dramatic than the backlash. Lozano told SFGATE the pictures were only meant to be temporary while the couple opened their first restaurant and cared for three children under the age of 5.

A Chicago deli offered essentially the same explanation. Chi-Town Deli manager Ali Malik told Business Insider that its AI-generated sandwich images were a stopgap while the restaurant waited for a new sign featuring actual photographs.

Before AI Slop, There Were Mashed Potatoes

For all its futuristic weirdness, fake food in advertising is a remarkably old practice. In the 1965 Supreme Court case FTC v. Colgate-Palmolive Co., the justices even weighed a familiar advertising trick: using mashed potatoes as a stand-in for ice cream.

The fake food itself was not necessarily the problem. Rebecca Tushnet, a Harvard Law School professor who specializes in advertising and trademark law, told Fortune that advertisers may use images that are not real as long as they do not mislead consumers about what they are buying.

"If a reasonable consumer would understand that the image was making actual claims about the composition or quality of the food, then that could be deceptive," she said. But if the image simply shows the general kind of thing a customer will get, "then it is not likely to be deceptive."

Generative AI makes that old distinction considerably stranger. Instead of mashed potatoes pretending to be ice cream, restaurants can now advertise food that never sat in front of a camera at all.

"The image itself can make a claim about the product's size, quantity, ingredients, quality, or appearance, even if the ad never says those things in words," Libby O'Neill, partner and deputy chair of advertising, marketing and promotions at Loeb & Loeb LLP, told Fortune.

And an "AI-generated" label is not necessarily a get-out-of-jail-free card. "It explains how the image was made; it does not correct a false impression about what is actually being sold," she said.

Her rule of thumb is simpler: "Sell the product you depict, and depict the product you sell."

For now, the question of where that line falls is being worked out in public, one slop-shaming Reddit thread and viral post at a time—while regulators, platforms, and restaurants alike adapt to imagery tools that are cheaper and faster than anything the makers of mashed-potato ice cream ever had.

This story was originally featured on Fortune.com.