AI has no taste — and that is your advantage

The New Yorker documented it in June: the "smooth, symmetrical sameness" pervading AI-generated interfaces. The same rounded corners. The same gradient blobs. The same purple-and-blue color pairs. The same pillow buttons. The same soft shadows on every surface.

When everyone builds with the same models and the same prompts, the output converges. I wrote about this in The vibe-code gap is real — the distance between a generated prototype and a production product is not shrinking. But there is another gap that matters more: the gap between generated output and memorable output.

That gap is taste. And it is the only advantage that compounds the more AI tools improve.

On this page

• What does it mean that AI has no taste?
• How AI sameness shows up in products
• What the slopless movement gets right
• Three things you can do about it today

What does it mean that AI has no taste?

Taste is the ability to know what to leave out. AI does not have it. Models predict the next token based on statistical patterns in training data. They produce the average of everything they have seen. That average is competent — clean layouts, correct structure, functional interactions. It is also interchangeable with every other output the same model produces for the same task.

Andrej Karpathy's term "vibe coding" described the experience of building by describing what you want and letting AI fill in the implementation. The problem is that the "vibe" is never yours. It is the aggregate vibe of the entire internet. Smashing Magazine featured a case study from MacPaw's design team building Eney, an AI assistant that deliberately rejects the default AI aesthetic — pink instead of blue, circular instead of rectangular, intentionally human-animated. Their conclusion: AI-generated design is technically correct but emotionally neutral. Breaking that neutrality requires specific, opinionated choices that the model would never make on its own.

How AI sameness shows up in products

Three patterns I see constantly:

Visual convergence. Every app uses the same stack: rounded cards, soft shadows, thick sans-serif headings, gradient accents. The colors vary but the structure is identical. Open five AI-built products side by side and cover the logos — you will struggle to tell them apart.

Functional conservatism. AI generates the safest possible interaction pattern for every use case. A settings page gets a list of toggles. A dashboard gets a sidebar and a grid of cards. A landing page gets a hero, features, pricing, CTA. Nothing wrong with any of these — but nothing specific either.

Copy that pleases everyone. Model-generated microcopy hedges. "Streamline your workflow." "Supercharge your productivity." The sentences are grammatically perfect and factually empty. No tension, no specific promise, no reason to care.

A Developers Conference keynote this year put it bluntly: velocity is not taste. When every builder has equal access to velocity, taste is what survives.

What the slopless movement gets right

The "slopless" movement crystallized around a simple diagnosis: AI-generated products wear a uniform. Purple gradients, thick sans-serifs, blobs, rounded corners stacked on more rounded corners, pillow buttons, soft shadows on every surface. The uniform is not bad — it is just not yours.

The correction is not to reject AI tools. That ship has sailed. It is to introduce a human editorial step between "generated" and "shipped" that injects intentional friction. Design CI tools now scan for AI signatures and fail builds that match model defaults. Designers embrace imperfection, grain, texture, and visible human authorship as marks of quality.

The insight is specific: the model produces the average. The winning product produces the exception.

Three things you can do about it today

Override the first output. The model's first draft of any UI is the most generic version of that UI. Instead of accepting it, prompt for a second version with constraints: "Make this feel less like a default template." The specific instruction forces the model away from the average.

Introduce one deliberate constraint. Pick something arbitrary and commit to it: no borders on cards, only one font weight, a single accent color used everywhere. Constraints force identity. The model never imposes them.

Review every interaction for tension. Read your microcopy aloud. Does it sound like a person wrote it, or does it sound like a committee decided it was inoffensive? If it is the latter, rewrite it to say something specific enough to be wrong. Specificity is the opposite of slop.

Taste is not a gift. It is a muscle you build by rejecting the default, one decision at a time. AI makes the default faster than ever. That only makes the people who override it more valuable.