Your flyer looks like ransomware.
I say that with affection, because I have shipped worse. If you have walked past a bulletin board, scrolled a local business page, or looked at a restaurant window in the past year, you have seen the genre: warped text, six-fingered mascots, gradients that seem to be sweating, and a headline font that is somehow both Papyrus and Impact. People are calling it a “ChatGPT flyer pandemic,” and the complaint that keeps surfacing is not really about ethics. It is that the posters are just ugly.
That distinction matters a lot if you build bots for a living.
Two separate arguments got glued together
There are two conversations happening at once, and mixing them makes both dumber.
The first is about provenance and fairness. When an AI-generated poster won the Ohio State Fair’s poster contest, Ohioans objected, and reporting noted the winning entry contained common AI-generated image errors. That is a legitimate fight about what a contest is for. A poster contest is a competition between people, and entering a generated image is a category error, like bringing a motorcycle to a footrace.
The second conversation is simply about craft. Readers sending in the worst examples they could find were not running provenance forensics. They were reacting to bad kerning and mangled hands. Those posters would be bad if a human had made them by hand at 2am.
The ethics argument is worth having. But if you are a builder, the craft argument is the one you can actually do something about tomorrow morning.
Why the output is bad, mechanically
Most terrible AI posters share a root cause: someone asked a single model to do four unrelated jobs in one shot. Copywriting, layout, typography, and illustration are different disciplines with different failure modes. Image models are genuinely good at one of those and genuinely bad at another. Asking for “a flyer for my taco truck with the text TACO TUESDAY 5PM in a fun font” hands the model a typesetting task it was never built for, and then we act surprised when it returns letter-shaped noise.
The fix is boring and architectural. Separate the concerns.
- Generate the image, not the poster. Ask the model for a background, a texture, an illustration, an isolated subject. No text in the prompt output at all.
- Do typography in a typography tool. SVG, HTML and CSS rendered to PNG, Canvas, Pillow, a design template with slots. Text you control is text that stays spelled correctly.
- Use a template layer. Grid, margins, hierarchy, a locked type scale. A flyer with three sizes of one good typeface and real whitespace will beat a generated masterpiece with garbled letters every single time.
- Add a check step. Run generated imagery past a vision model with a specific question: are hands correct, is anything in this image unreadable, does the subject match the description. Cheap gate, high catch rate.
That pipeline is a few hundred lines of glue code. It is also the difference between “this looks like a small business made an effort” and “this looks like a phishing attempt.”
The humans are winning, and that’s the useful signal
What I find most interesting in the current coverage is that designers are stepping in to repair these posters. Human designers are taking AI output and fixing it, and human creativity is being valued over the generated stuff. Framed as a contest, that is AI losing. Framed as a workflow, it is the actual answer sitting in plain sight.
Designers are not fixing these flyers by regenerating them with a better prompt. They are fixing them by applying judgment: cutting copy, choosing one typeface, aligning things, deciding what the poster is for. That judgment is the part your bot does not have. So build the bot to do the parts it is good at, and leave a clear seam where a person makes the calls.
What I’d actually build
If a client asked me for a poster bot today, I would not build a poster bot. I would build an asset bot plus a layout engine plus a human approval step. The model produces raw material and variations. The template enforces taste. A person picks, edits the headline, and ships. Fast, cheap, and it does not end up in someone’s “worst flyers” submission pile.
The people complaining about these posters are not anti-AI. They are anti-ugly. That is a solvable engineering problem, and treating it as one is more useful than arguing about whether the machine should have entered the contest.
🕒 Published: