# How to judge what AI makes for you

By Bart Dunweg, Jan 6, 2026 · 5 min read

An AI model will give you twelve competent options and an opinion about none of them. How I form my own opinion before I generate anything, and the three questions I ask before I accept what comes back.

Ask an AI model for a dashboard layout, the screen of charts and numbers a business checks every morning, and it will give you twelve. Each one is competent, none is obviously wrong, and all of them look as if they came from the same product. A single competent option used to take a junior designer an afternoon. Now the hard part has moved: generating options was never the scarce skill, choosing the right one is.

The model cannot do that choosing for you. What it hands you looks like a decision and usually is not one. Design systems, the shared building blocks a product is made from, taught me the difference years before AI did. A colour code only says what something looks like. A name like "the colour of the thing you press" says what it is for, and that is a decision. A generated screen, a draft of a text or a computer-made picture is the first kind: a description of what a plausible answer looks like for a request like this one. Whether it is right for this product, this user, this moment is a separate question, and the answer depends on things that are not in the request.

So the work splits in two. Before you generate anything, you need an opinion of your own. After, you need to check what came back against it.

## The gap between plausible and right

I saw the gap most clearly in something with nothing to do with software. I had an AI model make pictures of a house I had just bought, to see the rooms before a contractor touched anything. The pictures were consistently plausible. The proportions were confidently wrong until I supplied real measurements. The materials looked convincing and matched nothing sold in a hardware store. The light was always golden hour, because that photographs well, regardless of which way the room faces. A house does not negotiate. It is exactly the size it is.

The same gap shows up in interfaces: a screen that is consistent, on brand, usable by everyone, and still wrong, because it answered the letter of the request instead of the intent behind it. Nothing about that screen looks unfinished, which is exactly what makes it easy to wave through.

## Before: form an opinion first

The opinion you bring is what people call taste, and the word is often used as if it means what someone happens to like. It is closer to a compressed history: a record of everything you have looked at closely and formed a view on: the interfaces that felt cheap despite doing everything right on paper, the two typefaces that looked fine apart and fought on the page, the version of a brand that photographed well and meant nothing.

A model has seen more design than any person ever will and has an opinion about none of it, because it has never had to live with a decision, watch it fail, and carry that into the next one. Taste is judgment with scar tissue. A model has the shape of the judgment without the scars.

That is why the order matters. The tempting workflow is to generate first and react to what comes back, because reacting is easier than starting from nothing. But reacting to twelve options is a much shallower kind of judgment than forming a point of view before anything is generated. So I sketch a rough opinion first, on paper or in a sentence. The model's output then gets checked against something I thought, rather than becoming what I think because it arrived first and looked finished.

### The trap that looks like taste

There is one habit that passes for taste and is not: rejecting things because they do not look like what you would have made yourself. That is a mirror, not judgment, and it is worse than no filter at all, because it throws out good ideas that arrive from an unfamiliar direction along with the wrong ones. A model will generate options you would never have come up with, and some of them are better for exactly that reason. The question is whether something is right on its own terms, not whether it is how you would have done it.

## After: three questions before I accept anything

I ask the same three questions of AI output whatever it is: a screen, a paragraph of text, a picture of a room. I do not let myself skip to the last one.

### Would I defend this choice out loud, for a reason other than "it looked right"?

Almost everything a model produces now clears the bar of looking right, so that bar no longer filters anything. What still filters is whether there is a reason behind the choice that would survive being said to another person.

### Does it hold against the constraint that cannot move?

Every request has one or two non-negotiables in it, and a model treats every condition you give it as equally soft unless told otherwise. Most of them are soft. Finding the one that is not, and checking the output against that one specifically, is the part a model cannot do for itself.

### What would I change if I had built this from scratch, in the same time?

If the honest answer is nothing, the model did real work and I checked it, which is the good outcome. If the answer is a list, that list was the job, and treating the first draft as the finished one skips it.

## Where to spend the time a model saves

The instinct with a fast tool is to generate more. The better use of the time it saves is looking harder at fewer options, which was always the scarce part and is now the only scarce part left.

And taste is not built by watching. Looking at a thousand good interfaces builds a library, which is raw material for judgment but not judgment itself. Judgment gets built by making a call, seeing what happens when it is wrong, and adjusting, on things that are real, even small ones. A decision without consequences teaches nothing.

Everything mechanical around this, the drafting, the variations, the first pass at a hundred small choices, a model does well and I am glad to hand over. The choosing is not waiting to be automated. A model will always bring the twelve options. It will never bring the opinion, so the opinion has to be in the room first.