The first thing you notice when you start designing with AI is that it has answers. Endless answers, instantly, at any hour, delivered with total confidence. The second thing you notice is that most of them are wrong. Or almost wrong. Or (this is the one that gets you) right in a way that’s somehow still wrong.
So here’s the claim: the skill that separates good work from derivative slop right now isn’t prompting. Everybody’s prompts are fine. The skill is diagnosis. The model hands you something that looks finished, and the whole job compresses into a single question: what kind of wrong am I looking at?
The public conversation about AI design, meanwhile, is a shouting match between two camps. Camp one says AI is killing design: junior roles are toast, and the rest of us will end up as prompt typists chasing whatever aesthetic the model decided was current this week. Camp two says AI is democratizing design, anyone can do it now. (Anyone cannot.) Both takes share the same flaw, which is that neither has spent much time inside the actual work.
The four kinds
I have. Enough that the wrongness has started to sort itself into four kinds, and each one asks for a different move.
Broken-wrong is the cheap kind. The model misread the brief, lost the plot, invented a fifth brand color. No client will ever see it. It costs you a glance. Throw it away, go again, mourn nothing.
Almost-wrong is the useful kind. The thinking is right; the note is a little ‘off’. The composition holds but the tone is a hair too playful. The layout works but the type is doing too much. You don’t re-prompt your way out of almost (people try; it becomes a doom loop of increasingly desperate politeness: please… just slightly less playful). You put your hands on it and fix it. This is still the fastest route to good work I know.
Interestingly-wrong is the rare kind, and the fun one. It’s not what you asked for, and it’s better than what you asked for. A tonal accident. A strange visual tension. A composition you’d never have art directed on purpose, because no reasonable person would have. The model stumbled into it statistically; your job is recognizing there’s something in there worth stealing back from the machine.
And then there’s average-wrong, the dangerous kind. Nothing is broken. Everything is defensible. The work is plausible everywhere and decided nowhere, and it’ll sail through every review until somebody finally asks what it’s for. The problem is that the other three kinds announce themselves. This one just LOOKS finished. Most of the bad AI design shipping right now (and there’s a lot of it) isn’t broken-wrong. It’s average-wrong that sailed through six rounds of approvals wearing a lanyard.
Stop calling it a tool
Sorting the output is half the job. The other half is understanding why the machine deals you each kind, and that starts with retiring a metaphor. AI is not a tool. A pen is a tool. Photoshop is a tool: it does what you tell it, then waits, politely, forever. AI interprets. It anticipates, refuses, suggests, and occasionally gets confidently lost in a direction nobody asked for. Eric Rodenbeck calls it a medium rather than a tool, which is closer. My version: it’s a collaborator who’s read the entire internet, including the parts written by shit gibbons, and misunderstood a meaningful percentage of all of it.
Every designer I know who bounced off AI (myself included, for a while) bounced off it the same way: by treating it like a vendor. Brief in, deliverable out. Then the work comes back generic, or wrong, or eerily close to right but tonally off, and your first instinct is that the machine fucked it up. It did not. It did exactly what a system that interprets does: what it thought you meant… which is a very different contract than the one you believed you signed.
The dance
So the working relationship is a dance, and it has exactly two moves.
Leading means walking in with a point of view sharp enough that the model can’t hide behind generic answers. You hold the brand, the audience, the effect the work needs to have, the things you refuse to ship no matter how polished they look, and you keep narrowing the space the model gets to fill. The designers who lead well aren’t the ones writing the longest prompts; they’re the ones whose taste stays alive while they work.
Following means letting the process alter the brief. Interestingly-wrong only pays off if you’re willing to stop and ask whether the detour points somewhere better than the plan. The bad-designer instinct is to bin the unexpected because it wasn’t part of the original idea. (I’ve done this more times than I want to put in writing.)
The full move is both. Lead, watch what comes back, hold the line or chase the opening, lead again from the new position. What comes out of that loop is almost always better than what you walked in with, and… this is the part that keeps me here… almost always better than what the model would have made alone.
The machine is the consensus
That covers three of the four kinds. Average-wrong gets its own section, because averaging is not a flaw in the machine. Averaging is the machine.
Ted Chiang called ChatGPT “a blurry JPEG of the web” in The New Yorker: a compressed, slightly lossy copy of everything already made, details softening toward whatever surrounded them. The same pull lives in every AI design tool. If you ask it for something, you get a statistical guess at what something like that usually looks like. Sit with the word usually for a second. The model has no preference for good. It aims, every single time, at the center of what it has seen.
Good design has never once come from aiming at the statistical center. The point of design is to find what doesn’t exist yet, and next isn’t in the training data. This is why work made entirely by AI feels familiar in a way that eventually becomes exhausting: it resembles everything else because it is, quite literally, built from the average of everything else.
The model doesn’t push against consensus. It is the consensus, holding a pencil. The designer’s job is to push against that gravity, to treat the model’s fluency with the past as a launch point instead of a landing place.
In practice, the pushing is unglamorous. I’ve had rounds where ninety percent of the output was unusable and buried in one variation was an accidental idea I never would’ve reached on purpose. The ratio isn’t the point. The recognition is.
None of this threatens the core of the discipline, for the record. The floor of design output has risen dramatically, acceptable is now automated (I wrote about that shift in The Rising Floor), and what the rising floor clarifies is that nobody hires a designer for acceptable anymore. What is left is exactly the work this essay is about: holding one coherent instinct across hundreds of small decisions while everything pulls toward the middle.
Where the next eye comes from
Here is the part that genuinely worries me.
Taste used to be built on reps, and I’m old enough to have done mine. You looked at a thousand bad layouts before you could feel why a layout was bad. (Several hundred of them were mine.) The reps were never the point; the reps built the eye that eventually decided what the point was. The AI now does many of those reps for you. A junior designer with a model can produce work that looks senior almost immediately, and it’s often indistinguishable from every other piece made by someone using the same model on a similar brief. The junior can’t always tell. The model definitely can’t. The senior can, but the senior remembers what it cost to learn the difference.
The senior designers I know are worried about this. The junior designers I know mostly aren’t. Which feels like useful data on its own.
If the reps built taste, and the reps are becoming optional, where does the next generation of taste come from? Not from the model; it’s optimized toward consensus, and taste is the willingness to leave it. The most promising answer I’ve seen is to treat the output as the beginning of a harder design conversation instead of the conclusion of one. If the reps diminish, the discussion around the work has to deepen.
From the outside, AI design looks like a productivity story. From the inside, it’s a judgment story, wall to wall. The model can generate options, accelerate execution, and surface directions nobody asked for. It can’t decide what matters. It can’t tell polished from meaningful, and it can’t recognize the idea that feels risky before it feels inevitable.
Broken, almost, interesting, average. The model will keep dealing you all four, forever, without ever knowing which is which.
Knowing is the job.