AI-generated interfaces look generic because, without specific direction, a language model produces the most probable design for a request. The most probable landing page has a gradient hero, a centred headline, three feature cards and a familiar sans-serif. The fix isn’t better adjectives. It’s giving the model information it can’t guess: who the product is for, real references, firm constraints and explicit things to avoid.
The cause: models default to the centre
Anthropic’s applied AI team calls this distributional convergence. In their words, “Without direction, Claude samples from this high-probability center.” Patterns that appear most often in training data are the safest predictions, so they turn up whenever a prompt leaves room.
This explains why output from different tools looks so similar. They aren’t copying each other; they’re all converging on the same average. It also explains why the problem gets worse with vaguer prompts. “Make a modern SaaS landing page” describes thousands of existing pages, and the model returns their midpoint.
The recognisable defaults
Most people can now spot an unedited AI-generated interface. The usual tells fall into a few groups.
Colour
- Purple or indigo to blue gradients, often on white.
- An accent colour used everywhere, so nothing stands out.
- Low-contrast grey text for anything secondary.
Typography
- One common sans-serif at a handful of default sizes.
- Weak contrast between heading and body, or headings that are all the same weight.
- Body copy with a cramped line height or very long line length.
Layout and components
- A centred hero with a headline, a subhead and two buttons.
- Rows of identical rounded cards, each with an icon, a title and two lines of text.
- Even spacing everywhere, so groups don’t read as groups.
- Dashboards made of four stat cards above a chart above a table.
Content
- Placeholder-feeling copy: “Streamline your workflow”, “Built for teams”.
- Icons chosen because they exist, not because they mean something.
None of these are wrong in isolation. They’re generic because they’re chosen by default rather than by decision.
The fixes, layer by layer
1. Give it a point of view
The model needs a reason to leave the centre. Describe the product’s character in concrete, visual terms and commit to one direction. Anthropic’s guidance is to pick a clear aesthetic and avoid overused fonts and timid, evenly distributed palettes. If you can’t choose yet, explore a few directions deliberately before committing.
This is an internal tool for freight dispatchers who keep it open
all day. Direction: calm, dense and utilitarian, closer to a
trading terminal than a marketing site.
- Dark neutral background, one amber accent used only for status
that needs action.
- Monospaced numerals, 13px base size, tight 32px table rows.
- No gradients, no illustrations, no rounded cards. Hairline
dividers instead of boxes.2. Show, don’t describe
A reference image carries more information than any paragraph. Share two or three references and say what to take from each. Our guide to giving Claude Code visual references explains how to do this without the output becoming a copy.
3. Name the defaults you don’t want
Explicit avoid-lists work because they remove the most probable options. Anthropic notes that instructions such as avoiding particular default fonts improve results straight away. Keep the list short and specific to your product.
4. Constrain the system
A defined type scale, spacing scale and colour palette removes hundreds of small default choices. If your project has these as tokens, tell the agent to use only them. If it doesn’t, decide them before generating more screens.
5. Use real content
Generic copy produces generic layouts. Real product names, real data shapes and real edge cases (a 60-character company name, an empty account, a failed payment) force layout decisions that placeholder text never does.
6. Review and iterate in specifics
Most generated screens are close enough to refine. The guide to improving AI-generated UI without rebuilding everything covers a structured review pass.
Make the direction persistent
Every fix above is information about your taste and your product. Typing it into each prompt is slow and inconsistent. The lasting fix is to write it down once, in CLAUDE.md, a skill, or a tool your agent can query, so every session starts from your direction rather than the average. We go into this in how to give AI coding agents better design taste, and the difference between direction and rules in design references vs design systems.
Attuna is built around that idea. Your references, likes, dislikes and principles live in a Taste Profile, and your coding agent can request the parts that apply to the task in front of it.