You’ve probably felt it without naming it. A cafe menu where every bagel sandwich looks flawless, symmetrical, and a little too smooth. Something’s off, even if you can’t say what. TechCrunch AI reports that generative AI menus have landed in the restaurant business, and the reason they feel wrong points to a deeper problem in how these models work.
The short version: AI image models are converging on one narrow, “pleasing” aesthetic. And that sameness is starting to show up everywhere, not just on your lunch table.
What’s actually happening
Diffusion models and LLMs like Midjourney and ChatGPT learn by finding patterns in huge datasets, then predicting what you probably want. Ask for a fast food menu, and the model reaches for what it already knows: Wendy’s, Burger King, McDonald’s. Those menus already share a look. So the output copies that look, and the loop tightens.
Reality Defender CTO Alex Lisle put it bluntly to TechCrunch AI: “A lot of this stuff looks like a Chili’s menu from 2015, and there’s a reason for that. That was the corpus of work from which [the models] drew their function.”
Lisle draws a useful distinction. Model collapse is the extreme case, what he calls “almost like a mad cow disease” when a model feeds on its own outputs until it breaks. What we’re seeing with menus is milder. It’s convergence. The quality degrades, edges get sanded off, but the model still works. Every ice cream scoop turns perfectly round. Shrimp start eating their own tails.
Why it matters now
Here’s what stands out. This isn’t only an aesthetic complaint. People are reacting with genuine disgust.
Researchers at the University of Duisburg-Essen found that AI food images hit an “uncanny valley” effect, per TechCrunch AI. Images that looked almost real triggered more unease than images that were obviously fake. Lee Rainie, who directs the Imagining the Digital Future Center at Elon University, told TechCrunch AI that optimizing for “pleasingness” turns into homogenization. “What AI is known to do both in images and language is to shave off the edges.”
The editing habit makes it worse. One X user showed what happens when you generate a menu in ChatGPT, then edit it 100 times. The food drifts further from reality with each pass. TechCrunch AI replicated the experiment and got similar results. Restaurants tweaking prices and item names over and over are quietly running the same degradation loop.
The bigger signal
The convergence problem scales past food. If models keep training on their own smoothed-over outputs, the whole visual and language space narrows. That’s a structural risk for anyone building on top of these tools, not a quirk.
And it feeds a second industry. Reality Defender sells AI-detection and verification tools, a category growing precisely because trust in images is eroding. Lisle’s warning goes further than menus: “Seeing and hearing has always been believing, to the point where even our court systems are entirely tuned to the idea that the gold standard in evidence is taped confessions and videotaped evidence. That’s no longer the case.”
Practical takeaways
For restaurants and small businesses:
- Skip AI-generated food imagery. The backlash is real, and the uncanny valley effect means customers register unease even when they can’t explain it. Real photos win.
- If you must use AI, stop re-editing the same image. Each pass pushes it further from believable.
For AI practitioners and product teams:
- Watch your training pipeline for synthetic contamination. Convergence creeps in before collapse does, and it’s harder to spot.
- Diversify source data on purpose. If your outputs all trend toward one “safe” style, that’s a signal, not a feature.
- Build for provenance. Content verification is becoming table stakes, not a niche.
The menu weirdness is a small, visible symptom of something larger. Models trained to please keep shaving off the edges, and the more their own work leaks back into the pot, the flatter everything gets. You can read the full reporting at the original source.