---
date: Aug 30, 2026
author: Murat Alpay
role: Product Designer, Design Engineer
category: product
featured: true
cover: /images/equal-numbers-are-not-equal-brightness.webp
excerpt: Ask a model to make a dashboard feel premium and you are giving it permission to make a surprising number of decisions on your behalf.
---

# I’m asking AI to make fewer design decisions

When I first started generating interfaces with AI, I gave it a lot of adjectives.

Clean.

Premium.

Polished.

Modern.

Delightful.

It felt natural. Those are words designers use when describing an intention.

They are also almost useless constraints.

Ask a model to make a dashboard feel premium and you are giving it permission to make a surprising number of decisions on your behalf.

Now there is a gradient.

Cards begin floating inside other cards.

A settings screen gets a welcome hero.

The button gets an icon because buttons apparently get lonely.

Every hover needs movement.

Nothing is obviously terrible.

That is what makes this kind of output dangerous.

It is plausible.

## Adjectives hide decisions

I’ve gradually replaced most of those words with rules.

Not:

Make the dashboard cleaner.

More like:

There is one primary action on this page.

Do not nest cards.

Numbers align to the right.

Secondary metadata disappears before the primary value truncates.

Filters stay in the toolbar.

Do not introduce a second surface for an action that can happen here.

The output becomes less surprising.

I consider that a feature.

There are still many things I am happy for AI to decide.

It can build a loading skeleton.

It can wire up pagination.

It can make the fifth CRUD dialog.

It can choose between two spacing values that I genuinely do not care about.

But I’ve become much more protective of decisions that affect hierarchy.

What appears first.

What disappears first.

What deserves a label.

What becomes a new page.

What remains visible when the user cannot act.

Where the product interrupts someone.

These decisions are still expensive even when producing the interface is almost free.

## Better models did not reduce the need for intent

I expected the opposite.

I assumed that as models became better, I would need to specify less.

For implementation, that has mostly been true.

I describe the behavior and increasingly get working behavior back.

But in product design, leaving something unspecified does not remove the decision.

It delegates it.

The model still has to put something there.

So it reaches for a very reasonable average.

The average SaaS toolbar.

The average onboarding screen.

The average settings page.

The average empty state.

And the average is becoming extremely competent.

I think that is going to confuse a lot of teams.

The interface will look good enough that undefined product decisions stop looking undefined.

That is a different kind of slop.

Not ugly software.

Unconsidered software with very good typography.

## Generation made judgment more visible to me

There was a time when a lot of design quality was tangled up with execution.

Could you make the interaction?

Could you prototype it?

Could you build the component?

Could you make the page look coherent?

AI keeps making those things cheaper.

I like that.

But it also removes some of the places where weak decisions used to hide.

If I can create five viable versions of a page in an hour, producing another version is not particularly valuable.

Choosing why one of them should exist is.

I’m using AI more than ever.

I’m just asking it to make fewer decisions that I actually care about.