AI posters are bad for boring reasons.
Not mysterious reasons. Not “the models aren’t there yet” reasons. If you’ve spent any time on a bulletin board, a restaurant window, or a community group feed lately, you’ve seen the genre: warped text, six-fingered mascots, gradients that look like a screensaver from 2004. One outlet called it a “ChatGPT flyer pandemic” and asked readers to send in the worst examples. Readers delivered. And the verdict wasn’t really about AI versus not-AI. It was simpler. The posters are just ugly.
I build bots for a living, so my reaction to that wasn’t defensive. It was recognition. Ugly output is almost never a model problem. It’s a pipeline problem.
What the flyer pandemic actually reveals
Here’s the pattern I keep seeing in these posters. Somebody opened a chat window, typed “make me a poster for my bake sale,” and shipped the first image that came back. That’s it. That’s the entire system. One prompt, one generation, zero review.
Now compare that to how a designer works. They ask who the poster is for. They pick a typeface on purpose. They set a hierarchy so the date and time read from across a parking lot. They print a proof and squint at it. They throw out the first three versions. The work isn’t the rendering, it’s the judgment applied before and after the rendering.
When you strip a process down to a single generation step, you’ve deleted every stage where judgment lived. The model didn’t fail. The workflow around it never existed.
The Ohio State Fair poster contest made this uncomfortably concrete. An AI-generated entry won, and people noticed the winning poster contained common AI image errors. Sit with that for a second. The errors were visible enough for the public to spot after the fact, but nobody caught them before the ribbon went on. That’s not an AI scandal. That’s a missing review step in a human process.
The stages everyone skips
If I were shipping a poster generator, I wouldn’t start with prompt engineering. I’d start with the parts of the pipeline that have nothing to do with the model:
- Constraints before generation. Canvas size, safe margins, brand colors, one typeface family. A generator with no spec produces slop by default. Give it a spec and you’ve eliminated half the failure modes before the first token.
- Text out of the image. Diffusion models garble type. Stop asking them not to. Generate the artwork, then composite real text as a separate layer with real fonts. Boring engineering, dramatic quality jump.
- Automated checks. Contrast ratios you can compute. Reading order you can test. Whether the date string exists at all in the final render. Same instinct as a test suite, pointed at a visual artifact.
- A human gate. Not a rubber stamp. Someone whose job is to reject. If nothing ever gets rejected, you don’t have a review stage, you have a formality.
Notice how little of that list is about picking a better model. Most of the quality gap in generated design comes from stages that cost almost nothing to add and that almost nobody adds.
Why human designers keep winning
A designer has been going through these AI posters and fixing them, and the framing in the coverage was blunt: it’s AI versus humans, and the humans are winning. I don’t think that’s a permanent scoreboard, but I do think it’s an honest read of where things stand right now. Human creativity is still what people actually value in design, and the fixes prove why. A designer looking at a broken poster doesn’t ask for a new generation. They ask what the poster is trying to do and then make it do that.
That skill is the thing a lot of builders assume they can skip. You can’t. You can automate rendering. You cannot automate taste, and if you don’t have it, you need to route around your own absence by hiring it, borrowing it, or building enough structure into your system that it doesn’t need much.
What to take back to your own bot
This isn’t really about posters. Swap in code generation, copywriting, support replies, anything. The failure mode is identical: one call to a model, no spec, no validation, no gate, straight to production. The output looks plausible enough to ship and wrong enough to embarrass you later.
Generation is one stage. Treat it that way. Put a spec in front of it and a reviewer behind it, and your output stops looking like a ChatGPT flyer.
The models are fine. Our pipelines are lazy.
🕒 Published: