The tedious part of a junior designer's job was never only labour. It was the curriculum. The rough layouts, the fourteenth variation of a card, the screen for when there is nothing to show yet, the screen for when something went wrong, the version for a name that is far too long: that work was slow and unglamorous, and it was where people learnt what a good decision looks like. AI now does most of it, and does it well. Teams have noticed the speed. Fewer have noticed that the training went out of the door with the tedium.
I co-founded a design studio in 2014. The reps are the endless repetitions that trained a designer the way practice trains an athlete. This piece is about what replaces them, because something has to, and it will not happen by itself.
The reps were teaching something
Drawing every state of a screen by hand does not teach drawing. It teaches noticing. Somewhere around the sixth state, you discover the case nobody specified: the list with one item, the account with no history, the payment that is neither approved nor declined. You discover it because you had to put something in the box, and you could not put something in the box without deciding what it was.
That is the first rep of judgment: being forced to choose because the work does not move until you do. The second rep came afterwards, when a senior looked at what you chose and asked why. The third came later still, when the thing shipped and you saw what people did with it.
A model collapses the first rep. It fills every box with something plausible, and the junior's job becomes picking from the results. Picking is a real skill, and I have argued elsewhere that taste gets built by deciding rather than by watching. The trouble for someone at the start of a career is that they are now asked to pick before they have done enough deciding to know what they are picking for.
Supervising what you never made
There is an old name for this trap. Researchers at Purdue looked at how design practitioners talk about AI in their work, in a short paper on the ironies of AI-assisted design (opens in a new tab) at CHI 2025, the main research conference on how people use computers, and they borrowed it from Lisanne Bainbridge's 1983 paper on the ironies of automation: when a system takes over the hands-on task, the person is left to monitor it, and the hands-on skill that good monitoring depends on stops being practised. The authors are careful to say they are reporting what practitioners worry about rather than measuring skill loss, and they note, citing earlier work, that the concern falls particularly on junior designers. Their phrase for the risk is that "excessive delegation may unintentionally hinder skill development".
That is a fair summary of the problem without overselling it. Nobody has shown yet that juniors trained alongside models end up worse. What is clear is the shape of the risk. A senior reviewing AI output draws on years of having made that output by hand. A junior reviewing the same output has nothing underneath the review except the output itself.
How judgment still gets built
The reps can be replaced. They cannot be skipped. These are the four I would build a junior's week around.
Write the decision before the generation
Before a junior asks a model for anything, they write down what they expect a good answer to be and why, in a few sentences. Not as a ritual: as the thing that gets reviewed. A senior who reads the intent first and the output second can see whether the junior is thinking or selecting, and can teach the difference.
Defend the choice out loud
A model cannot be cross-examined about why it made something. A junior can, and should be. Why this option, what did the rejected ones get wrong, what would make you change your mind. The answers will be weak at first. Weak answers given out loud are a rep; strong-looking output chosen in silence is not.
Review against a stated intent
Reviewing is now the main act of design for most people on a team, so it has to be practised like a craft. Give a junior a stated intent and a set of outputs, some generated and some made by hand, and ask which ones meet the intent and where the others fail. The skill being trained is the one the paper above worries about: telling a confident answer from a correct one.
Put real stakes on it, even small ones
Judgment without consequence is opinion. Let a junior own something that ships, an error message, a settings page, a single flow, and let them see what happens: the support questions, the numbers, the colleague who builds it and asks what happens when the field is empty. Small stakes still teach, because they are real.
What a senior has to do differently
Seniors used to teach by marking up a junior's designs. That worked because the design was the junior's thinking made visible: every line had been drawn on purpose, so correcting a line corrected a thought. A generated design carries no such information. Marking it up teaches the junior how to prompt better, which is useful, and very little about design.
So the review has to move from the design to the reasoning. Ask for the intent before opening the file. Ask what was rejected. Ask what the junior would have done without the model, and take the answer seriously, because it is the only part of the work that is theirs.
The harder change is personal. A senior with a model can fix a junior's work in less time than it takes to explain what is wrong with it. That has always been true to a degree, but it is now true almost every time, and the temptation to fix rather than teach grows with it. Every time it wins, the team gets the right screen and loses the lesson.
Seniors also have to narrate their own judgment more than they used to. Juniors once saw a senior's discarded sketches lying around and could work out what had been rejected. When the rejected options live and die inside a chat with a model, the reasoning is invisible unless someone says it.
Faster and shallower
The risk is not that teams get worse. It is that they get faster and shallower at the same time, and the first hides the second for a long while.
A team that has stopped training judgment still ships good work, because its seniors are still carrying the judgment. The output looks the same. The speed looks better. Nothing in a quarterly review shows that the next generation is learning to supervise work it has never made. Depth shows up years later, when the seniors move on and the people who replace them turn out to know how to choose between options and not how to tell when all of them are wrong.
The reps were never the point. The person they produced was the point, and now that the tools do the reps, someone has to produce that person on purpose.