The best AI feature is the one that changes less

The result might look polished. But now you have to check everything.

This is one of the most frustrating experiences with AI design tools: asking for a specific adjustment and getting a reinterpretation of the whole thing.

Sometimes, you already like what you have. You just need help with one detail.

Canva's announcement of Brush caught my attention for that reason. It describes a model that works with an existing design, understands its structure, and makes requested changes while keeping the elements editable. Canva also acknowledges something essential: the model needs to understand what should stay exactly as it is.

Knowing what to preserve is part of understanding the request.

Think about how we give feedback to another person. "Can you make this heading shorter?" carries an expectation about scope. We expect a shorter heading. If the person also changes the colors and moves the image, we have to stop and explain which decisions were already settled.

With AI, that conversation can become a loop.

You ask it to restore the image. It restores the image but changes the spacing. You fix the spacing, then notice that the wording has shifted again.

Each correction becomes another opportunity for something else to move.

The effort after the edit

The experience continues after the AI finishes generating.

There is the moment when you compare versions. The uncertainty about whether something changed outside the area you were watching. The decision to keep going or return to an earlier version.

That effort deserves more attention.

A tool can complete an action quickly and still leave you with a slow review process. If you have to inspect the entire design after every request, it becomes difficult to work with confidence.

Predictability matters: I should have a reasonable idea of what will happen when I ask for a change.

That does not mean every action must be tiny. A request to update all headings could affect an entire presentation. The scope is broad, but the intention is clear.

The frustration starts when the scope expands without your direction.

Make the change understandable

A good editing experience should make it easy to answer three questions:

• What changed?
• Does it match what I asked for?
• How do I get back to the previous version?

A clear comparison can help. So can a short summary, the ability to accept individual changes, and an undo action that restores the expected state.

That information also needs to work beyond visual highlighting. Someone using a screen reader needs an accessible way to understand and review the result. W3C guidance describes how status messages can announce action results without moving focus, while warning against excessive announcements that interrupt the experience.

"Three headings updated" is a useful starting point. From there, the person should be able to inspect those headings and reverse the change.

Confidence comes from repeated interactions

One impressive result can make someone try a tool. Consistent, understandable behavior gives them a reason to keep using it.

When a tool respects the scope of a request, you learn how to work with it. You can make an adjustment, review it, and continue without wondering what else happened.

That is the experience I want AI design tools to get better at.

There is still plenty of room for bold exploration and unexpected ideas. But when I ask for a precise edit, precision is the experience.