Designing for the Apprentice

There's a kind of work nobody misses. Forty variants of the same button. Screen-by-screen specs so engineering wouldn't have to guess. Empty states, errors, the edge cases the main design hadn't accounted for. For years it was the entry-level work of the craft: what you gave whoever had just arrived, because it was necessary, because it was boring, and because someone had to do it. And it's exactly the work AI now does better and faster.

What almost nobody said at the time is that this work also taught. Not because anyone designed it as training, but by accident. Making the thirtieth variant of a button forced you to notice why the twenty-ninth didn't work. Writing the error state forced you to think about what happens to a person when something goes wrong. Fixing a spec a senior had handed back marked in red was, without anyone calling it that, a lesson. The learning came disguised as production. And now that production is being automated, the disguise falls away, and the learning falls with it.

What the numbers show

Dolphia Arnstein gathered some data worth looking at carefully. According to the analysis she cites, entry-level design hiring at large tech companies has dropped by half since 2019 and now represents around seven percent of new hires. These are figures from one sector, not the whole discipline, but the direction is clear. And her most interesting observation isn't the drop but a paradox: the tools sold as a way to democratize design demand more judgment, not less. If you can't tell a result that works from one that only looks like it does, the tool doesn't help you: it lets you fail faster. The floor didn't come down. It went up.

That's the knot. The machine does the junior work, but evaluating that work requires senior judgment. And senior judgment was formed, in large part, by doing the junior work. Takuma Kakehi framed it as a law of automation: work is allocated according to what the machine can do, never according to what a person would benefit from learning. Nobody decided that training designers had stopped mattering. We simply automated the part that trained them.

An argument with my own argument

This topic makes me more uncomfortable than others, because it touches two ideas I've been defending.

In Designing for Pressure I argued that some friction isn't an obstacle but the mechanism through which understanding forms, and that a tool that removes every demand works like an environment with no pressure: comfortable, and slowly weakening. If that holds for someone who already has judgment, it holds far more for someone who doesn't yet. Whoever has gone through the friction can delegate from a formed judgment. Whoever starts out delegating has nowhere to delegate from.

In Designing for Inherited Judgment I wrote that judgment doesn't travel as a rule but as a case, and that what passes from one generation of a team to the next can be a set of conclusions or the practice that produced them. Something new appears here. More and more teams are writing their judgment into rubrics for agents to apply, and those rubrics work: Richa Sehgal showed how a review that used to take forty-five minutes came down to six. But think about whoever joins that team. What they find isn't the discussion that produced the rubric, but the finished rubric. They learn to apply it, not to think it. They inherit the conclusion without having been in the conversation.

And there's a third piece that complicates everything. Kai Wong observed that a high-fidelity screen said something it never said out loud: that a competent person had made a hundred decisions there. When making it took two weeks, that time was a signal of the judgment invested. Now that it's generated in minutes, the screen no longer vouches for anyone. This doesn't only affect how an experienced designer is evaluated. It also affects how a junior with a good eye gets recognized, because there's no longer a piece of work in which that eye becomes visible.

The positions in dispute

Not everyone sees this as a problem, and it's worth listening to those who don't.

Maya Brennan, for example, proposes something radical: close Figma for a week and start building directly with AI. In her reading, the traditional learning path shouldn't be compressed but skipped, because the skills it taught are losing value. It's a serious position. It assumes judgment can be formed another way, by shipping real things and watching how they fail, instead of spending years making variants nobody uses.

Slava Polonski goes another way. He asks which jobs should remain human, and proposes that the answer shouldn't depend on whether the machine can do them, but on whether the human being is part of what we value. One of his tests is precisely formation: there are tasks worth having a person do because doing them forms that person, even if a machine could do them better. It's a way of allocating work by its value, not by the machine's capability.

And Nicole Michaelis takes the discussion to taste. She describes it as a muscle that only grows under resistance, and argues we've stopped feeding it: less exposure to what we don't expect, more copying of what already works. I like her image, but it's worth saying that the decline of taste, in her text, is a premise, not something measured.

The risk of nostalgia

There's an objection I raise against myself and don't want to dodge. It's easy to romanticize the friction that formed us. The tedious work also wore people down, kept them for years on tasks that taught them nothing new, and often depended on being lucky enough that a senior would take the time to correct you. Not everything that cost us served us. Henry Desroches puts it very well while talking about something else: we can't go back, among other reasons because we remember it wrong.

So the question can't be how to preserve the learning we had. That friction was accidental, and the accidental isn't preserved: it gets replaced by something deliberate, or it gets lost. The question is what formative friction we design now, knowing it will no longer come free with production.

Designing formation

If the friction that forms no longer shows up on its own, it has to be designed. And here a criterion that emerged in discussions of accessibility helps me: friction is productive when it's chosen or co-designed by whoever has to go through it. Imposed, it looks more like a barrier or an arbitrary rite of passage. Chosen, it becomes training.

Some concrete forms appear almost on their own when you look at the problem this way. Having someone do by hand, once, what they're later going to delegate, so they know what they're delegating. Having them compare their attempt with the machine's and explain how they differ, instead of just approving the result. Making reviews hand back the reasoning and not just the correction, because judgment travels as a case with its why. Presenting a team's rubrics as a starting point for discussion rather than a rule to follow. None of these practices is new. What's new is that they no longer happen by accident.

And there's an underlying problem no individual practice solves. For decades, the formation of designers was a free byproduct of production: you learned by doing work that had to be done anyway. When a machine does that work, formation becomes an explicit cost. Peter Merholz pointed out that organizations only fund what they can see and anticipate, and the formation of judgment is among the hardest things to see and anticipate.

Maybe that's the question left open for those who design teams, not just screens. If judgment is no longer formed as a side effect of the work, someone will have to decide to form it on purpose. Who is going to pay for that, and what will happen in ten years to the teams that didn't?