Which Kind of Wrong

The first thing you notice when you start designing with AI is that it has answers. The second thing you notice is that most of them are wrong. Or almost wrong. Or right in a way that is somehow still wrong.

The skill that decides whether you produce good work or derivative slop is not how good your prompts are. It is whether you can pinpoint which kind of wrong you are looking at.

That skill does not have a name yet.

The conversation about AI in design has been dominated by two camps that mostly talk past each other and past the actual work. The first camp says AI is killing design. Generative tools will hollow out junior roles and leave the rest of us as prompt typists chasing whatever aesthetic the model decided was current this week. The second camp says AI is democratizing design. Now anyone can be a designer. The barriers are gone. Welcome the citizens.

Both takes miss the same thing, which is the actual texture of doing the work.

I have spent enough hours inside that texture to know the model does not replace the designer. It also does not democratize the discipline. What it does is hand you something that looks finished and ask you to know whether it actually is.

And it is not really a prompting problem. It is closer to a dance.

To get to what I mean by that, it helps first to kill the metaphor that keeps getting in the way. AI is not a tool in the traditional sense. Calling it one flattens what actually happens in the room.

Photoshop is a tool. A pen is a tool. A tool does exactly what you tell it to do and then waits for the next instruction. AI behaves differently. It interprets, anticipates, refuses, suggests, and occasionally gets confidently lost in a direction you never asked for.

That interaction is less like operating software and more like collaborating with a partner who has read the entire internet and misunderstood parts of it. Eric Rodenbeck has written about this from inside the academy, framing AI as a medium rather than a tool.

I have struggled with AI whenever I treated it like a conventional tool. So has almost every designer I know who tried that approach first. We write prompts the way we would write briefs to vendors, expect deliverables on the other end, and then wonder why the output comes back generic, or wrong, or weirdly close to right but tonally off.

The surprise is usually the first clue that the interaction is not as mechanical as we pretend it is.

A traditional tool would have done what you asked. A system that interprets does what it thinks you meant.

The dance has two moves. Leading and following. Both matter, and the designer who can only do one of them usually produces bad work in different but predictable ways.

Leading means walking into the conversation with a point of view. Not a complete answer, but an opinion sharp enough that the AI cannot hide behind generic responses. You hold the brand in your head, and the audience, and the emotional effect the work needs to create, and the things you refuse to ship no matter how polished they look.

The AI does not know any of this until you tell it, and even then it only understands a flattened version.

Your job while leading is to keep refining the shape of what you want, narrowing the space the model gets to fill in. The designers I know who lead well are not the ones writing the longest prompts. They are the ones whose taste stays alive while they work.

Following requires a willingness to let the process alter your original assumptions.

The AI will occasionally hand you something you did not ask for that points toward a better version of the brief than the one you started with. It might be a tonal instinct, or a composition, or a strange visual tension that solves a problem you had not even named yet.

I have had moments where ninety percent of the output was unusable, but buried inside one variation was an accidental idea I would never have art directed intentionally. The model stumbled into it statistically. My job was recognizing there was something there worth stealing back from the machine.

The instinct of a bad designer is to throw unexpected output away because it was not part of the original plan. I have done that more times than I want to admit. The instinct of a good designer is to pause and ask whether the detour is actually pointing somewhere better.

This dynamic rarely gets discussed directly.

The AI pushes back, and the pushing back is sometimes the most useful thing in the room.

The full move is doing both. Lead, watch what comes back, decide whether to hold the line or follow the new opening, then lead again from the new position. Repeat.

The work that comes out of that loop is almost always better than what you walked in with, and almost always better than what the AI would have produced on its own.

Neither side gets there alone.

To understand what only the designer can do, it helps to understand what the AI literally cannot.

Large language models, and the image models behind most AI design tools, learn patterns from enormous amounts of training data. They are built to predict what is most likely to come next given everything they have already seen.

This is a remarkable trick and a profound limitation at the same time.

Ted Chiang described the limitation memorably in The New Yorker in 2023 when he called ChatGPT “a blurry JPEG of the web.” A compressed, slightly lossy version of everything already written. The image holds together when you stand back. Up close, the details soften toward the average of what surrounded them.

The same gravitational pull exists in AI design tools.

When you ask an AI to make something, what you get back is a statistical guess at what something like that usually looks like. ‘Usually’ is the important word there. The output drifts toward the mean of everything in the training data.

Good design has never come from aiming at the statistical center.

The point of design is to find what does not yet exist, or to pull something into shape that has not been shaped before. That is a forward-tense activity. AI is fundamentally past-tense. It can show you what has been done. It cannot tell you what deserves to be done next, because next is not 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 constructed from the average of everything else. The model has no preference for what is genuinely good. It cannot distinguish novelty from mimicry.

The model does not push against consensus.

It is the consensus.

The designer’s job is to push against that gravity. To use the AI’s fluency with the past as a launching point instead of a landing place. To recognize when the model has handed you the average and decide whether the work needs something the average cannot provide.

This is the part both the doomer and utopian narratives miss.

AI does not threaten the core of design because design was never about producing the average. AI does not democratize design because access to the average is not the same thing as being a designer.

Giving someone access to PowerPoint does not automatically make them a presentation designer. Giving someone access to a camera does not automatically make them a cinematographer. The tools matter, but judgment matters more.

The discipline of good design has always advanced from the edges inward. The tools evolve faster than the underlying responsibility does.

There is a version of this story that is actually good news, and the industry has been strangely reluctant to tell it.

The floor of design output has risen dramatically. Things that once required real technical skill can now be produced by almost anyone with a prompt and a few minutes. A passable landing page. A clean slide deck. A logo that does not embarrass the company. A visual system that looks competent enough to survive a client meeting.

None of it will win awards, but none of it will lose accounts either.

The instinct is to read this as a threat, and in some ways it is. Work that took years to learn can now be produced in minutes by people who never had to learn it. That loss is real.

For those of us still doing the work, the rising floor is also clarifying.

The work designers are actually valuable for becomes more visible precisely because mediocre execution stops being scarce. Nobody hires a designer anymore to merely produce acceptable work. Acceptable is now automated.

What still matters is the strategic and aesthetic core of the work: the part that depends on someone holding a coherent instinct across hundreds of tiny decisions while resisting the gravitational pull toward sameness.

That was always the valuable part. It was just buried beneath production labor.

The Photoshop revolution hinted at this shift on a smaller scale. Before Photoshop, huge portions of a designer’s day went into physically making things look polished. After Photoshop, production accelerated, and the conversation about what deserved to be polished became more important.

AI is doing the same thing again, just at a much larger scale.

If you are a designer who knows what you want, the work has never been easier to make. If you relied on the production grind to feel valuable, the rising floor is going to feel like the floor disappearing underneath you.

What has changed most dramatically is not just the speed of production, but the accessibility of fidelity. Ideas that once required entire teams or weeks of engineering support to communicate can now be explored interactively in an afternoon. Design reviews no longer have to rely on static screens and imagined transitions. A single designer can prototype movement, responsiveness, narrative flow, and interaction patterns at a level that used to sit much farther downstream in the process.

That shift matters. Not because the AI is making the decisions, but because it allows more of the actual thinking to become visible earlier. The distance between concept and experience has narrowed considerably.

Both experiences are real.

So far this sounds mostly optimistic, and mostly it is. But there is one part of the picture that genuinely worries me, and the industry still has not figured out what to do with it.

Taste used to come from doing the reps, and I am old enough to have done mine.

You looked at a thousand bad layouts before you could feel why a layout was bad. You produced a thousand mediocre logos before you could reliably distinguish mediocre from good. The reps were never the point, but the reps built the eye that eventually decided what the point was.

Taste was built through repetition, exposure, and correction.

The AI now does many of those reps for you.

A junior designer working with a model can produce work that looks senior almost immediately. The output is polished. The composition is defensible. It is also often indistinguishable from every other piece of work made by someone using the same model on a similar brief.

The junior designer cannot always tell.

The model definitely cannot tell.

The senior designer can, but the senior also remembers what it cost to learn the difference.

That tension is still unresolved.

If the reps were how taste was built, and the reps are becoming optional, where does the next generation of taste come from?

The model cannot teach it. The model is optimized toward consensus, and taste requires the willingness to depart from consensus. Someone still has to develop that instinct, and the old path for developing it is narrowing.

I am not sure the industry has a satisfying answer yet.

The senior designers I know are worried about this. The junior designers I know mostly are not, which feels like useful data on its own.

The most promising approach I have seen is using AI to accelerate production while treating the output as the beginning of a harder design conversation rather than the conclusion of one. If the reps diminish, the discussion around the work has to deepen.

The designers who will define what design means in five years are learning this interaction now. They are not avoiding the model, and they are not surrendering to it either. They are learning when to direct, when to listen, and how to recognize the difference between something that is technically correct and something that actually matters.

From the outside, AI design looks like a productivity story. From the inside, it is a judgment story.

The designer who enters the process with strong instincts and an open ear is doing something the model alone cannot do. The designer who expects the model to make the decisions is slowly training themselves out of the job.

I am not worried about design itself.

I am watching the taste question with real concern, and I think the industry should be more honest about it than it has been. Design has always depended on people willing to make the forward call, to recognize possibilities the consensus would never arrive at on its own.

A model cannot make that kind of judgment.

It can only approximate what has already existed.

That distinction matters more than most people realize.

AI can generate options, accelerate execution, and surface unexpected directions. But it cannot decide what matters. It cannot recognize the idea that feels risky before it feels inevitable. It cannot tell the difference between something polished and something meaningful.

Someone still has to make those judgments.

And the designers who learn how to work inside this tension, directing when necessary, listening when useful, resisting the pull toward sameness, are going to make things our predecessors could not have made alone.

The responsibility for deciding what matters never left the designer.

The model can approximate the past.

The designer still has to decide what deserves to exist next.