AI built a design system that looked right and was all wrong

We asked AI to build a design system for one of our apps. It opened the token file, picked a teal as the primary color, layered the shades nicely, named the tokens cleanly. It looked exactly like a good design system should.

The only problem was that it was wrong.

That color came from nowhere. The AI guessed it because it "looked plausible." When we flagged it, it switched to a blue, this time a real color that actually exists in the system, but still wrong, because it belonged to a different product, not the one we were working on. The correct color, a terracotta, only came out after we opened the real source file and read it. Three rounds, three colors that looked convincing, all wrong.

This is not a story about AI being dumb. It is a deeper trap, and the lesson changed how we let AI touch a design system for good.

Part 1Why "looks right" fools us

A good design system has a clear signature: layered colors, systematic token names, consistent spacing. AI is very good at this. It imitates the shape of a good design system almost perfectly.

But a design system is not judged by its shape. It is judged by whether it matches what is real. Our terracotta is not right because it is pretty. It is right because it is the color the live code uses this very second. The blue AI guessed was layered just as nicely, but nothing anchored it to the truth.

This is the point where AI invents over reality without us noticing, because the result looks too good to question.

Part 2Reflect, not invent

The rule we set afterward is short. Before AI touches any token, it has to go find where the real thing lives first: open the theme file, open the config the app actually uses, read the real value. No guessing from memory, no picking because it looks nice.

It sounds small, but it flips the whole mindset. Instead of asking AI to "design" a design system, we ask it to "reflect" the design system that already exists. Its job is not to be creative. Its job is to read what is real, accurately.

And there is a second layer that is even sneakier.

Part 3Know the family first, then reuse

The blue AI guessed on the second round was a real color, one that genuinely exists in the system, but it belonged to another product. Once you have several products under one roof, each with its own design system, picking a real color is not enough. You have to pick the right one.

So we added one more guardrail before building. Every time we are about to touch a given app's screen, we have to answer first: which family does this app belong to, does that family's design system already exist, and only then reuse the right one, instead of building a new one every time we cannot recall.

What actually saved us was that when we were unsure, we asked the product owner directly instead of guessing. And the best answer we got back was "I don't know, don't guess it for me, go open the real thing." Admitting you don't know beats making up an answer that sounds confident.

Part 4Put it to work

If you are having AI help you build a design system, here are the short rules you can use right away.

1. Reflect before invent. Have AI find the existing token file first, always. Found it, use it. Not there, then create.
2. Every value needs a source. Any color or value the AI adds must point to which file it came from. If it can't point, it guessed.
3. Check the family before you reuse. Know which family the app belongs to, then pick the right one.
4. When unsure, ask, don't make it up. With people, and with the AI itself.

"Looks right" is the most dangerous trap in design work with AI, because it slips past your eyes so easily. The way to guard against it is to force everything to have a source before you trust what you see.