Over the past year I’ve looked at a great many things made entirely with AI by people who don’t have a design background. Web pages, decks, PDFs, the occasional dashboard. Most of them are things their makers could not have produced a year earlier, and that’s worth pausing on before anything else. Someone who has never written a line of HTML now has a working page. That matters.
But I’ve noticed something about my reactions to them. My notes are the same notes. What is this meant to be? Why does everything here carry the same weight? Where do you want me to look first? What should I do next? I have been writing those comments in the margins of other people’s work for more than twenty years. For most of that time, the other people were designers.
If my critique of what AI makes is indistinguishable from my critique of what trained professionals make, that tells me something important.
The AI vernacular
People who spend time with AI-generated writing learn to spot it quickly. There are the verbal tics — the things that “quietly” fail, the obsession with digging “moats,” and the reflexive construction where something is declared to be not one thing but another. That last one sits on the list of phrases I would have found bothersome before AI but somehow almost never read. Now, it’s everywhere.
Design has a vernacular too, and it’s just as easy to spot. Anthropic said as much last November, in a post about improving Claude’s front-end output: because “safe design choices — those that work universally and offend no one — dominate web training data,” models “converge toward generic, ‘on distribution’ outputs,” producing what users had started calling the AI slop aesthetic. The same post told the model to stop reaching for Inter and “purple gradients on white backgrounds.” (I mourn that one a little bit, because I love Inter almost as much as I love the em dash.) Adam Wathan, who created Tailwind, had already apologized that summer for making every button in Tailwind UI indigo five years earlier, “leading to every AI generated UI on earth also being indigo.”
What happened next is the instructive part. Steered away from purple, the output found a new center. The current version of Anthropic’s design guidance warns instead against a warm cream background, a high-contrast serif and a terracotta accent — the accent Anthropic uses in Claude itself, which the guidance now calls “a tell.” Close one door and the model walks through the next most probable one. The vernacular is simply what probable looks like at a given moment.
What probable looks like
If more than two decades working on the web has taught me anything, it’s that probable has never been good. Most web pages don’t do their job. They don’t make good use of the few seconds of attention a visitor gives them; they don’t answer the questions a person asks the instant a page loads. That was true long before any of this, and it was true of pages made by designers. The fundamentals — hierarchy, contrast, proximity, rhythm, symmetry used for a reason — are taught unreliably, when they’re taught at all. You never really know what a given designer learned. And then there’s taste: most web pages are ugly. But that’s another post.
So a model trained on the web learned the web’s average, and the web’s average was already failing.
Which means something most designers wouldn’t happily admit: AI has reached the median. Whatever the median designer could produce, a person with no training can now produce in minutes, for close to nothing. If your work lived at the median, the threat is very real.
I suppose the good news is that being better than the median was always the job. It’s just that until recently, reaching the median cost something.
The same notes
That’s why the sameness of my notes reads to me less as a warning than as a clarification. The critique hasn’t changed because the problem hasn’t changed. Production is now nearly free, and what remains as hard as ever is knowing what a page is for and what should happen on it. That knowledge is what the fundamentals are, and it’s why they remain the most valuable thing a designer can own.
The obvious response is to teach the machine the fundamentals. People are trying. I’ve tried. Why that keeps falling short is a question for another essay. But my bet is that no matter how good the machine gets, it won’t relieve the human of the responsibility for review, critique, and decision-making.