Interfaces vs. intelligence: designing with AI as a partner, not an intern
Most teams adopted generative AI as automation. A cheaper, faster intern that writes the first draft, fills the mockup and summarizes the research. That works for chores. For the creative part of design, the research says it can quietly make your work worse.
This article looks at what controlled studies found about AI and design fixation, why AI-assisted work tends to converge, and how to set up a workflow where AI widens your options instead of narrowing them. It's for design leads and designers who already use AI daily and want better ideas from it as well as faster ones.
§ 01
Faster first drafts, narrower ideas
The speed gain from AI is real and easy to see. The cost is harder to see, because it shows up as the ideas you never had.
At CHI 2024, researchers from the University of Melbourne ran a controlled experiment with 60 participants on a visual ideation task. Half had an AI image generator to help. The result, published as The Effects of Generative AI on Design Fixation and Divergent Thinking, was blunt. The AI group fixated more on the initial example and produced fewer ideas, with less variety and lower originality than people working without it.
The researchers also noticed that outcomes depended on how people used the tool. Participants who treated the generator as an answer machine anchored hardest. That detail matters more than the headline, because it means the problem is the workflow, not the technology.
§ 02
Why AI pulls everyone toward the same answer
A second study shows the same effect at the level of a whole field. In Science Advances, Anil Doshi and Oliver Hauser asked people to write short stories, some with ideas from a language model. Stories written with AI help were rated more creative and better written, especially for weaker writers. They were also more similar to each other. Every writer improved, and the collection as a whole got narrower.
Design has the same dynamic. Models are trained on what already exists, so their first suggestion leans toward the average of everything they've seen. When every team starts from that first suggestion, product categories start to look alike. Checkout flows, onboarding screens and landing pages converge on one competent, forgettable pattern.
For an individual designer that's a fair trade on a deadline. For a company trying to stand out, it's a slow loss. The fix isn't to stop using AI. It's to stop accepting its first answer as the starting point.
§ 03
Four jobs AI does well in a design process
It helps to name the jobs where AI widens your options, and keep it there.
Synthesis comes first. Summarizing interview notes, clustering survey answers and pulling themes out of support tickets saves days, and the designer still decides what matters.
Next is the blank canvas. A few rough directions get a team talking faster than an empty file. Treat them as conversation starters, never as drafts.
Then volume. Ask for twenty variations, not three, and ask for them in deliberately different directions: different layouts, different metaphors, different constraints. Variety is the point, so judge the batch by how different the options are from each other.
A prompt that does this well names the axes on purpose. For a pricing page, ask for one version built around a calculator, one around a comparison table, one that leads with a customer story and one that hides the price until the user picks a use case. You'll throw most of them away. The one you keep will sit further from the category default than anything a plain "design a pricing page" prompt would give you.
Last, fidelity on demand. Turning a sketch into a clickable flow or a rough flow into a polished screen lets you test the right question at the right level of detail.
What these have in common is that the human frames the problem and makes the call. AI covers ground faster.
§ 04
Connect AI to your design system, not the internet
A lot of generic AI output comes from generic context. A model that has never seen your tokens, components or accessibility rules will invent its own, and you'll spend the time you saved fixing them.
The Model Context Protocol is an open standard for giving AI tools structured access to your data and systems. Figma's own MCP server uses it to pass variables, components and design context from Figma files into coding tools like VS Code, Cursor and Claude Code. Generated code then uses your real color, spacing and type tokens instead of guessing.
Accessibility belongs in that context too. If the model knows your contrast pairs, focus styles and minimum target sizes, its output can follow WCAG 2.2 from the first draft instead of failing an audit later. Spell out the rules you care about, such as 4.5:1 contrast for body text and visible focus on every control, and check generated screens against them like any other review item.
The same idea applies to research and brand. Give the model your guidelines, past research and the component inventory, and ask it to cite which ones it used. Output grounded in your own system is easier to review and much harder to mistake for everyone else's product.
§ 05
A workflow that keeps the ideas yours
Here's how I'd set it up, based on what the fixation research suggests.
Sketch before you prompt. Spend ten minutes on your own rough ideas before opening any AI tool. It gives you an anchor that's yours, so the model has to compete with it.
Ask for spread, then prune. Request many options in explicitly different directions, cluster them, and throw out the ones that look like the category average.
Keep decisions human. Research synthesis can be automated, variant generation can be assisted, but final copy, critical flows and sign-off stay with a person who can explain the choice.
Measure more than speed. Time to first prototype is easy to track and tells you little on its own. Also track how many distinct directions a team explored, and whether the shipped design performed better in testing. If AI made you faster and every test came back flat, it wasn't helping.
The designers who get the most from AI won't be the fastest generators. They'll be the ones who know when to ignore the first answer.