Visual design is where AI feels most impressive. Modern tools produce clean colours, balanced spacing and neat components in seconds.
That speed changes the job. Making a screen look good is now cheap. Making sure it works is not.
The happy-path trap
AI loves ideal conditions. Short names. Neat lists of three. Clean cards with perfect photos.
Real products are not like that. Real products are full of:
- Empty states, when there is no data yet.
- Loading states.
- Network errors and timeouts.
- Permission errors.
- A table with 10,000 rows instead of five.
- A name that is 40 characters long.
AI designs the happy path beautifully. Products are mostly the other paths.
If your work stops at a polished happy path, that part is now easy to automate. A good-looking screen is no longer the finish line. It is the starting point.
AI as a reviewer
Vision models, the kind that can look at an image, are surprisingly good at reviewing layouts. But the quality of the review depends on what you give them.
Without context: you upload a screenshot and ask “how is this layout?” You get generic tips about contrast and white space.
With context: you give the screen plus the user’s main goal, the technical limits and what matters most on this page. Now it can flag real problems:
- A secondary button that looks more important than the main one.
- A path where the eye jumps around the screen in a messy way.
- Related settings split across menus that do not connect.
Same tool. Very different result. The difference is the context.
From static mockups to code
The old way was to draw rectangles in a design tool and hand them to an engineer to rebuild.
That gap is shrinking. AI can turn a layout idea into working front-end code. That means you can test real states in a browser instead of drawing dozens of static frames for hover, error and long translated text.
The craft moves from pushing pixels to setting up design tokens, layout rules and how things behave.
Prediction is cheap. Checking is the job.
However good the screen looks, the engine underneath has not changed. It is still predicting what a screen like this usually looks like.
- It does not know if the design works. It has no idea about your users, their workflows or your rules.
- It cannot promise accessibility. It can style a form cleanly, but it cannot guarantee keyboard use, screen reader order or legal compliance.
- The checking is yours. You test the empty state. You feed it messy data. You try the edge cases. And you are the one responsible when it ships.
The one line to remember
AI makes drawing the surface fast and cheap. Structuring the information, handling the edge cases and checking every screen against reality is still your job.
Checking matters even more when the AI starts acting on its own. That is next: the agent that lied to me.