Why does AI make everything look like corn?

Once you see it, you can't unsee it.

AI-generated food is everywhere now. Menus, delivery apps, the specials board at the place that definitely didn't hire a photographer. And a lot of it has the same weird texture. Smooth. Glossy. Little rounded nubs, evenly spaced, every one catching the light.

Corn.

Ground beef is the worst offender. Real ground beef is a mess. Crumbly bits, browned edges, a little rendered fat pooling in the pan, no two pieces alike. AI ground beef looks like someone sculpted taco meat out of really gross kernels. Close enough to pass at a glance. Wrong enough that your gag reflex notices before your brain does.

I couldn't let it go. So I went down the rabbit hole on why.

Why it happens

A quick disclaimer. I couldn't find anyone who has formally studied "the corn effect." I guess this makes me a pioneer. What follows is how these models work under the hood, and every piece of it points in the same direction. It's not one bug. It's five things stacked on top of each other.

1. The models are built to average

Image generators start with static, like an old TV between channels. Then they clean it up step by step, each time asking the same question: what's the most likely version of this picture?

That works great for the big stuff. Plate, taco, table, nice lighting. It works badly for detail that is random by nature.

Ground beef looks the way it does because of a process. It got ground, seared, broken up with a spatula, left to sit in its own fat. The model doesn't know any of that happened. It has never cooked anything. It only knows what bumpy brown food tends to look like in photos.

And its best guess at "bumpy" is a field of similar, rounded, evenly spaced nubs. Which is a pretty good description of corn.

The average of a thousand messy things isn't messy. It's smooth. Every piece gets rounded off toward the same shape.

Food isn't the only victim, either. Hair, grass and gravel all fail the same way. Food is just where we notice it most.

2. The fine detail gets thrown away, then made up

Most of these models don't actually work on the full picture. To save computing power, they work on a compressed version, roughly eight times smaller in each direction. At the very end, a second piece of the system blows it back up to full size.

The tiny irregularities that make meat look like meat are exactly what gets lost in that compression. So the system fills them back in with generic, plausible texture. And generic texture tends to come out tidy and repetitive.

Think of a photocopy of a photocopy, where the copier is confident enough to redraw the parts it can't read.

3. The training data was already fake

These models learned what food looks like from food photography. And food photography is styled.

Food stylists brush oil on burgers so they shine. They pick the most uniform pieces. They light everything so it glistens. Some of the most famous food shots in advertising aren't even edible.

So the model learned food from pictures that were already a little unreal, then exaggerated it. That's where the waxy, glossy, every-piece-catches-the-light look comes from.

4. Tuning for pretty makes it worse

After the base training, most image models get a second round of tuning toward images people rate as attractive.

People reward saturated. Glossy. Clean. So the models learn to push food toward a hyper-appetizing look that tips right over into uncanny. Same thing happens with faces. Over-filter a selfie and at some point it stops looking like skin.

5. The prompt is part of it, just a smaller part

Nobody describes texture. The captions the models trained on say "beef tacos," not "crumbly, uneven, browned edges, rendered fat." And the restaurant owner generating a menu image types "beef tacos" too.

So texture gets left entirely up to the model. And the model, as we've covered, defaults to the average.

You can push back. Words like "crumbled," "irregular," "charred bits," "harsh natural light" or "amateur phone photo" help. So does turning down the setting that controls how hard the model sticks to your prompt, which tends to over-polish things when it's cranked up. It gets better. It doesn't fully go away.

Why food is where you notice it

There's a sixth piece, and it isn't about the AI at all. It's about you.

Your brain is very, very good at judging food. Telling fresh from off was a survival skill long before it was an aesthetic one. Nobody taught you that meat that looks too smooth is suspicious. You just know.

A slightly wrong texture on a chair, a cloud or a sweater barely registers. On ground beef, it trips an alarm somewhere in the back of your head. That's the uncanny valley. Not wrong enough to be obviously fake. Just wrong enough that something feels off and you can't say what.

This isn't a food problem

Everything above applies to anything irregular. Food is just where it's easiest to see.

Read enough AI-written LinkedIn posts and you get the same feeling you get from AI ground beef. Same shape every time. Same three-part list. Same uplifting kicker at the end. Nothing wrong with any one sentence. The whole thing just reads like corn.

Same with strategy decks. Customer emails. Product positioning that could belong to any company in the category. The models are doing exactly what they were built to do. They're giving you the most likely version.

The most likely version of a strategy is the one everybody else already has.

The bumpy parts are the parts that came from somewhere. The deal you lost and why. The customer who used your product in a way nobody planned for. The opinion half your team disagrees with. The thing you tried in 2011 that didn't work and what it taught you.

That's the ground beef. It's irregular because it came out of a real process. And it's exactly what the averaging smooths away.

This matters most for the organizations I spend my time with. Careful, regulated, been burned before. Their edge was never being average. It's the weird, specific, hard-won stuff they know about their customers that nobody else does. Run that through a tool built to find the most likely answer and you don't get a sharper version of yourself. You get corn.

Bring your own texture

I'm not telling anyone to stop using AI. I use it every day. It's great at the plate, the table and the lighting.

The texture is your job. The specific story. The number you actually saw. The opinion you'd defend across a table. Put that in before the tool touches it, and check that it's still there when the tool is done.

If it could have come from anyone, it's corn.

Real ground beef is ugly. That's how you know it's real.