A while ago I asked an AI agent to take the section dividers off this website. The labels, the alignment and the whitespace were already separating the content. The lines were just sitting there.
Once they were gone, the spacing had to change. The first version put 80px between sections on desktop, which felt cramped, so I kept pushing it until 128px looked right. The agent changed the code and we wrote the decision into DESIGN.md.
None of that trained anything. The agent followed instructions, and nothing about it shows the model would make the same call in another project. Still, a vague preference ended up with something concrete attached to it: a layout, an alternative, and my correction.
A surprising amount of what we call taste can be talked about this way. Once I can point at a result and say why I prefer it, part of my judgment is out in the open, and someone else can pick it up.
So I'm uneasy when people say taste will keep designers safe from AI. I expect AI to eventually make judgments we'd fairly call good taste. I have no idea when. I just don't want my plans to depend on it never happening.
Taste can change
By taste I don't mean rules about fonts or a fondness for minimalism. I mean choosing what fits a particular piece of work, noticing when something doesn't, and changing it. Part of that is knowing when a perfectly attractive decision is wrong for the job.
The 128px gap is one choice in one composition. Turn it into a universal spacing rule and you've lost the point. A system that had actually learned my taste wouldn't just remember the number. It would notice when the reasons behind it stopped applying.
That's a lot harder than following a style guide. I think of it as an open research problem, though, not a line we've proven machines can't cross.
People are already working on pieces of it. ImageReward trains a reward model on human ratings and comparisons of generated images. Pick-a-Pic collects comparisons from users and uses them to train PickScore, which predicts which image people will prefer. Both came out in 2023.
Picking the better image is a much smaller task than designing a product. Guessing which picture someone clicks tells you nothing about whether a model can judge an unfamiliar interface after months of use. What these papers do show is that some preferences can be represented and learned. How far that goes, nobody knows yet.
I'll go further than the papers do. I think judgments about context and coherence will get more learnable as the systems and the feedback get better. That's my expectation. The papers don't prove it.
When good work gets easier to come by
Say I'm right. Someone with little design training describes what they need and works with a system that makes, compares and revises decent options. They still have to steer it toward something useful, but they need a lot less professional help to get there.
For people trying to make things, that's a real win. A small project could get a level of care its budget would never have paid for. More ideas could become usable before anyone has the money to hire a team.
For people who sell that work, like me, it's less comfortable. If a client can get an acceptable result while paying for less of my time, the years I spent learning the craft won't hold the old price up. The work can stay good and still stop being scarce.
Judgment is in the same position. It's tempting to hand production to AI and keep selection for ourselves: the model makes a hundred options and we pick the right one. But choosing between options is exactly the kind of thing a system might get better at. I wouldn't count on "final selector" being a safe place to stand.
I also can't promise that every task that goes away gets replaced by a better one. I don't know that. Some people will get more room to make things. Others will lose work they enjoy and rely on. Both can happen at once.
What I'd rather do is find out what these tools can actually take on, including the parts of my work I care about most. Declaring those parts impossible to automate would just be a way of not looking.
Whose taste gets repeated?
If AI can learn taste, it matters whose taste it learns.
Picture a system that gets rewarded when its images win quick side-by-side votes. A composition that reads instantly will do well there. A design that only shows its strengths after weeks of use needs a completely different test. The feedback we pick shapes the judgment we end up teaching.
That's part of why I don't equate popular with good. A preference can be widespread and still wrong for a given product. The quieter option might lose the side-by-side and be much easier to live with.
I ran into a version of this in my piece on DESIGN.md: a reference file carries the assumptions of the product it came from. A model's defaults deserve the same suspicion. If lots of people accept the same defaults, I'd guess certain looks get even more common. I don't think that's baked into AI, but I'd keep an eye on it.
Personalization might pull the other way. A system could learn one person's preferences, or help a project keep its character. That interests me far more than a universal score for how good something looks. The catch is that a tool that only learns what I already like could make it easier for me to repeat myself.
I want to keep running into work I don't like at first, trying directions that feel foreign, and changing my mind. A good design tool should leave room for that instead of treating every past preference as a standing order.
Working without a guarantee
For now, I can get better at saying why I make the choices I make. When I write down why a reference fits a project, or why one version feels cramped, I leave myself something to argue with later. The tools I work with also get more to go on than "make it look nice."
Not every judgment fits in a document. I can start with the parts I can explain and pay attention to the ones I can't. I want to know whether a tool can carry a correction into a new context, when it needs more information, and whether its suggestions ever show me an option I'd have missed.
None of this makes my job safe. I prefer it because it keeps me working with what the tools can really do, instead of leaning on a limitation I can't prove is permanent.
I still care about taste. I just don't think it has to belong only to humans.