How AI Is Reshaping Creative Design

The question usually asked is whether a model can design. It is the wrong question, and it produces an argument neither side can win. The more useful one is narrower: which parts of this work were always judgement, and which parts were only the labour that judgement had to pass through to reach the page?

Split the work along that line and the last few years stop looking like either a threat or a revolution. They look like a change in where the expensive part sits.

What actually got cheaper

Most of what has changed sits firmly on the labour side. Producing twenty directions instead of three. Filling a layout with plausible copy before the real copy exists. Standing up a working prototype on the same afternoon the idea arrived. Turning a rough sketch into something a stakeholder can react to without a day of production in between.

None of that was ever the creative act. It was the toll paid to get to it. Removing a toll does not make the destination worth less — it makes the road cheaper, and that changes what is worth attempting at all.

The second-order effect matters more than the first. When exploring a direction costs an hour instead of a week, teams explore directions they would previously have talked themselves out of. The work gets braver at the edges, because being wrong stopped being expensive.

Speed is the visible change. Taste is the constraint.

Speed is the part most easily mistaken for the point. A team that can generate forty options and a team that can choose well between three are not the same team, and only one of those capabilities got cheaper.

Taste — knowing which of the forty is worth defending, and being able to say why — has quietly become the binding constraint on almost every project. It was always necessary. It used to be rationed by how many options anybody could physically produce, and that rationing has gone.

Which is why “we can produce more” is a weak claim on its own. Volume is only useful to a team that can also discard quickly and for stated reasons. Without that, more options is simply a larger pile to be indecisive about, and the indecision now happens later, with more sunk cost attached.

Where it genuinely helps

Research and synthesis

Grouping a thousand support tickets into themes is genuinely faster. So is summarising interviews, clustering feedback and getting a first read on a corpus nobody has time to read end to end.

What it does not do is decide which theme is the business's actual problem. That judgement depends on strategy, cost and appetite — things a model has no stake in. Treat synthesis as a way of reaching the shortlist faster, not as a way of skipping the argument about which item on it matters.

Iteration

A model will produce another variant indefinitely and never once tell you to stop. That is genuinely useful for widening a search early, and genuinely dangerous later, because “try another” is always available and always feels like progress.

Knowing when to stop is a judgement about the audience, the constraint and the risk. It is most of the job, and it is the first thing to get skipped when the next variant is one click away.

Prototyping

The gap between a static design and something you can actually click has narrowed considerably. That is a real change in how early an idea can be tested against reality rather than against opinion, and it moves decisions earlier — to the point where they are still cheap to revisit.

Where it does not help

Anywhere the cost was judgement rather than labour. Deciding what a business should be known for. Choosing which of three defensible directions fits a market you have actually looked at. Weighing whether a pattern some people find slightly confusing is still worth keeping because every alternative is worse somewhere else.

There is a related trap in interface work: designing as though the model will be right. That deserves its own treatment, and has one in designing for models that are wrong.

Authorship, and how it gets lost

Authorship survives all of this, but only where it is actually exercised. The failure worth worrying about is not designers being replaced. It is designers accepting the first output because it is competent — when competent is the floor these tools now guarantee, and the entire reason to hire anybody is what sits above it.

A generated starting point is a draft with nobody's name on it. It becomes work when someone decides what it is for, removes the parts that serve nothing, and takes responsibility for what is left. That decision is the thing being paid for, and it does not become less valuable because the raw material arrived faster.

The practical risk is homogeneity. Tools trained on what already exists are, by construction, good at producing more of what already exists. If a brand's job is to be distinguishable, the default output is precisely the wrong place to stop.

The floor rose. The ceiling did not.

The most durable change is what happened to competence. Work that is merely competent — a clean layout, a serviceable illustration, copy that reads properly — used to be a reasonable thing to sell, because producing it reliably was difficult. It is now close to free, and anything close to free stops being a differentiator.

What that leaves is everything competence was standing in front of: whether the thing is distinctive, whether it is aimed at somebody specific, whether it is making a claim worth making. Those were always the harder questions. They were simply easier to avoid when producing the artefact absorbed most of the budget.

Craft has not stopped mattering, but where it matters has narrowed. It matters most where an audience can actually perceive it — typography they read for minutes at a time, motion they sit through repeatedly, the density of a screen they use every day. It matters least in the places it was most often spent: producing variations nobody outside the team will ever see.

What it changes about briefs and reviews

Two practical things shift. Reviews stop being about whether the work is well made, because it usually is, and become about whether it is the right one. That is a harder conversation and a more useful one, and teams that have not prepared for it tend to fill the gap with taste-based notes that nobody can act on.

Briefs have to name the decision more sharply, because options are now cheap enough that a vague brief will generate forty plausible answers instead of three, and none of them will be wrong enough to reject cleanly. A brief that says what has to be true for the work to succeed is the only reliable defence against that — which is the argument in the brief is the deliverable, made more urgent by the tooling.

A working rule

Use the tools where the cost was labour. Decline to use them where the cost was judgement. Let them widen the search and shorten the loop, and do not let them make the choice.

In practice that means being explicit about which mode you are in. Widening — generate freely, cheaply, without attachment. Narrowing — put the generator down and argue from the constraint, the audience and the evidence. Most of the poor outcomes we see come from running both modes at once, where every difficult decision can be deferred by producing one more option.

It is how we approach AI design work, and how we tend to advise teams building AI products: the model is a fast collaborator with no stake in the result, which makes it excellent at proposing and unqualified at deciding.