AI tools are making it faster than ever to build an interface, and faster than ever to build the exact same interface everyone else is building.
Here are some thoughts about why that’s happening, why it’s a real business problem, and what actually stops it (spoiler alert, it’s a design system.) A lot of AI-assisted products are starting to look related these days, not in a ‘shared industry conventions’ way, but in a specific, almost identical way.
Why AI-Assisted Products All Look Alike
AI models produce a statistical average of the interfaces they were trained on. Not a bad interface, just an unoriginal one. When you ask a model to build you a dashboard, it isn’t drawing on a point of view. It’s predicting the most probable design, which is another way of saying the most common one.
That tendency gets worse the more AI is used. A 2024 study published in Nature found that when AI models are trained again and again on content that earlier AI models generated, they gradually lose their rarer, more unusual patterns and drift toward a narrower, more homogenized average. Researchers call this “model collapse.” This study wasn’t about interface design specifically, but the same mechanism applies to AI-generated UI: as more of the web fills up with AI-assisted interfaces, and those interfaces get used to train the next generation of models, and those interfaces get used to train the next generation of models, each new generation has less variety to draw from than the last.
There’s also a simpler, more mechanical reason. A lot of AI-generated frontend code leans on the same popular component libraries and starter templates, and many teams never customize past the defaults. If a popular library’s default button color or corner radius shows up often enough in the code these models were trained on, that default becomes the average, and the average becomes what gets generated for the next team, and the team after that.
None of this is new, either. A 2021 study presented at the ACM CHI conference (PDF) analyzed more than 227,000 screenshots from over 10,000 websites spanning 17 years and found that the visual similarity between websites’ layouts declined 44 percent between 2010 and 2019, meaning designs became dramatically less distinguishable from one another over that decade. The researchers traced the shift to two forces: the rise of mobile-first responsive design after the iPhone, and the dominance of a small number of frameworks and component libraries, like Bootstrap.
This time, the difference is speed. The convergence referenced above took most of a decade to play out, as frameworks and design trends slowly spread across the industry. AI skips that slow spread entirely. The bias is already baked into the model. One prompt is enough for the resulting design to come out looking like everyone else’s. Multiply that across every team using the same handful of tools, and a decade-long drift collapses into months, then reinforces itself every time a new model trains on the last one’s output.
Why It’s a Business Problem, Not Just an Aesthetic One
There’s a real cost hiding under what looks like a simple design nitpick.
When most AI-assisted products converge on the same look, visual presentation stops differentiating and becomes background noise. That pushes the competition entirely onto price and feature lists, which is a much harder way to win. A generic, templated feel doesn’t just fail to stand out; it can quietly undercut trust. An interface that looks like it was assembled from the same kit as a dozen other tools can read as unfinished or even second-rate, even when the product underneath it is solid.
This is similar to a problem we’ve written about before in the context of AI risk generally: something can look completely finished and professional while very little real thinking or intention happened underneath it. Here, it just shows up in pixels instead of a report.
The Fix Isn’t Avoiding AI. It’s Giving It Something to Push Against.
What actually matters is what you hand AI before it starts generating, not whether you use it for design and development.
The teams whose AI-assisted work doesn’t come out looking like everyone else’s aren’t avoiding these tools. They’re feeding them constraints. A design system with documented tokens, real components, and a specific point of view can keep an AI-assisted product distinct from every other AI-assisted product: it gives the model something to build inside of, instead of defaulting to the median.
A few ways to put that into practice:
Document the reasoning, not just the values.
A token file that lists a hex code or a spacing scale tells an AI tool what to use, not why it matters. Documenting the intent behind a choice (why this palette, why this much whitespace) gives the AI something to reason from in a situation the system didn’t explicitly cover, instead of falling back on its own default assumption.Show components assembled, not just isolated.
A component library on its own shows a button, a card, an input field, one at a time. Feeding an AI tool documentation of how those pieces actually combine in real screens gives it a much stronger signal for ideating an on-brand layout than isolated component specs do.Translate brand personality into concrete rules.
Words like “approachable” or “confident” mean nothing specific to an AI model; it has no way to render an adjective. If those words are supposed to shape the product, the system needs to spell out what they actually look like: specific imagery choices, motion behavior, a spacing philosophy, so the AI has something concrete to ideate from instead of the most statistically common interpretation of “confident.”Document what’s off-limits, not just what’s encouraged.
A system that only shows the ideal version leaves an AI free to fill every gap with its own default. Naming the patterns your brand deliberately avoids gives it a boundary to generate within, closing off the overused options it would otherwise reach for.Build deliberately unique elements into the system itself.
A custom interaction, an icon set, or an illustration style no common library shares creates a throughline across your whole product. When an AI tool draws from the system, that same element carries through automatically, giving its output a point of view the model’s training data doesn’t otherwise have.
A design system...gives the model something to build inside of, instead of defaulting to the median.
Bottom Line
AI didn’t create the sameness problem. Design has drifted toward a dominant look before, more than once, and that has been measured as recently as the 2010s. What AI changed is the speed: a decade-long drift can now collapse into months, reinforcing itself every time a new model trains on the last one’s output.
The fix is making sure whatever you hand these tools already has a baked-in point of view, not slowing down or stepping back from them. A documented, well-defined design system was always a foundation for consistency. Now it also keeps a product from looking like everyone else’s.
