What if the reason AI posters look so bad has almost nothing to do with the model generating them?
The current round of mockery is well earned. Readers have been sending publications their worst examples of AI-generated flyers, the ones spreading across social feeds, bulletin boards, and restaurant windows. One writer described the situation as a “ChatGPT flyer pandemic.” A poster made with AI won the Ohio State Fair’s poster contest and drew public complaints, partly because the winning image carried the familiar visual errors that give these things away. And now human designers are stepping in to repair the damage, which has been framed as humans versus AI, with humans winning.
I want to push back on that framing, not because I think the posters are good, but because I think the diagnosis is wrong. We are not watching a contest between human taste and machine incompetence. We are watching what happens when a generative model is deployed with no surrounding system.
A single forward pass is not a design process
Consider what a working designer actually does. They read a brief. They ask what the poster is for, who reads it, and from how far away. They choose a grid. They set type, then look at it, then hate it, then set it again. They check whether the headline still reads at three meters. They send a draft to someone who tells them the kerning is off. Then they fix it.
That is a loop with memory, an explicit specification, and an evaluation step that can reject output. Almost none of the flyers currently being ridiculed were produced by anything like that. They were produced by one prompt, one generation, and one human who accepted the first result because they had no basis for rejecting it.
So the failure mode is not “the model cannot design.” The failure mode is that a generator with no critic will always regress to the statistical middle of its training data, and the statistical middle of poster design is mush. Diffusion models do not know that a poster has a job. They know what posters tend to look like.
Why the errors cluster where they do
The tells that gave away the state fair winner are instructive, because they are not random. They concentrate in exactly the places where pixel-space generation is structurally mismatched to the task:
- Text. Typography is a discrete, symbolic system with hard correctness conditions. A letter is either an “R” or it is not. Painting it as pixels means approximating a symbol that should have been placed, not drawn.
- Hierarchy. A poster’s information order is a decision about meaning. It cannot be inferred from texture statistics.
- Counting and structure. Hands, fingers, repeated elements, alignment. These require a persistent representation of the scene rather than a locally plausible one.
- Negative space. Emptiness that does work is an intentional choice, and there is nothing in the objective function that rewards restraint.
Every one of these is a known weakness with known architectural answers. None of them require a better image model.
What a competent pipeline would look like
If I were designing an agent system for this, I would not let the image model near the type at all. Split the problem by representation:
Separate the spec from the render
Produce a structured layout first. Grid, regions, hierarchy, palette, type sizes, as explicit data. That artifact is inspectable, editable, and version-controlled. Then render imagery into the regions the spec allows, and place text as real text through a vector or layout engine that cannot hallucinate a glyph.
Add a critic that can say no
A generator without an evaluator is a slot machine. The evaluator does not have to be a model with taste. It can check measurable constraints: contrast ratios, minimum legible type size at the intended viewing distance, bleed and safe margins, whether every required piece of information is present. Most of the flyers being passed around fail checks this dull.
Close the loop and keep the trace
Revision is where design happens. An agent that generates, evaluates, and regenerates against the same spec produces something far better than one that generates once, and it leaves a record of why each choice was made. That record is what makes human intervention cheap instead of a rescue mission.
The designers fixing these posters are the missing component
This is the part I find genuinely useful about the current moment. When a human designer takes an ugly AI flyer and repairs it, they are performing the evaluation and revision steps that were never built. They are the critic, the spec, and the loop, operating manually and after the fact.
That work is real, and the judgment behind it is the scarce thing here. It also tells us precisely what these systems lack. The contest controversy is not evidence that A It is evidence that we handed people a raw generator, called it a design tool, and are now surprised that output quality tracks the sophistication of the process rather than the size of the model.
Better posters are available. They just require building a system instead of typing a sentence.
🕒 Published: