AI building tools now produce interfaces that look finished in seconds, and that speed hides a structural bias. Models trained on what already exists reproduce the most common patterns, not the ones that fit a particular audience or goal. Nielsen Norman Group researchers describe the result as output that feels "good from afar, but far from good."

That gap is where the human comes in. Taste, in the sense that matters here, isn't an innate eye for aesthetics; it's the practice of making intentional choices toward a vision instead of accepting whatever the averages serve up. Generation got cheap, but discernment didn't, so judgment now separates okay results from exceptional ones.

This lesson closes Module 3 by exploring why AI defaults toward the generic, how common-but-wrong patterns sneak into builds, when accepting a convention is actually the right call, and who holds the design judgment when a person prompts an interface into existence.

What is directing taste?

Directing taste means supplying the one thing an AI building tool structurally can't: judgment about what's right for this specific context, audience, and goal. Nielsen Norman Group researchers found that the pattern-matching behind these tools "drives them toward the most common solutions, not the most contextually meaningful ones" [1]. The tool knows what's frequent across the internet; only the builder knows what's right for the project at hand.

That makes the builder's role intent ownership, not decoration. Ethan Mollick describes the job in three parts: "you have to know what you want to create; be able to judge whether the results are good or bad; and give appropriate feedback" [2]. None of those steps can be delegated to the model, because each one depends on a vision the model doesn't have.

The stakes are simple: generation got cheap, and discernment didn't. When anyone can produce a working interface in minutes, judgment about which interface is worth shipping becomes the differentiator.

Good from afar, far from good

Good from afar, far from good

In late 2025, NN/g researchers ran a real design project through multiple AI prototyping tools and compared the results to a human designer's version. Even with detailed prompts, the outputs showed "a lack of visual hierarchy or grouping among related elements, overused colors creating visual tension, poor color contrast, inconsistent margin spacing" [3]. Each issue was small; together they separated a design that feels thoughtful from one that feels almost there.

One annotated example makes it concrete. Bolt placed a course-materials password and the link it unlocked far apart on the page, a grouping decision no rule forbids and nearly every experienced designer makes automatically. The tools didn't fail loudly; they failed in the details.

The researchers' summary gives this practice its name: the output "often feels good from afar, but far from good." Their paradox follows directly. AI lowers barriers to creation, yet it also magnifies the gap between okay results and exceptional ones.

Taste as intentionality

Design critic Elizabeth Goodspeed, quoted approvingly by NN/g, defines taste as "what enables designers to navigate the vast sea of possibilities that technology and global connectivity afford, and to then select and combine these elements in ways that, ideally, result in interesting, unique work" [4]. NN/g sharpens the definition further: taste is intentionality. Designers exhibit taste by being deliberate in the choices they make, which turns taste into a practice rather than a gift.

That reframing dismantles a common belief, the idea that taste is innate and a person either has it or doesn't. Intentional decision-making toward a vision can be learned, and the essay NN/g endorses is literally about how to acquire it.

The same article offers a useful analogy: anyone with a modern smartphone can take a technically high-quality photo, but without an eye for composition the photos won't be compelling. As Sarah Gibbons and Kate Moran put it, "Technical capability does not equal creative ability."

The generic default

Left without a style direction, AI building tools converge on the same look. NN/g's evaluation found that outputs defaulted to "a similar, generic look using sans-serif typeface and minimalistic styling," with many generated screens appearing flat and interchangeable [5]. The mechanism is structural: models trained on existing design reflect its most frequent patterns, and many tools lean on the same component libraries and frameworks, such as Shadcn and Tailwind CSS.

The sameness shows up at product scale too. When the researchers placed dashboards generated by Bolt and Claude side by side, both shared clean, understated styling with little visual differentiation or brand expression.

That's why the generic default isn't a flaw awaiting a patch. It's a property of training on what already exists, so the push toward a specific identity has to come from the person directing the tool. In a world where anyone can generate an interface, generic is the starting point and distinctiveness is a choice someone has to make.

Pro Tip! Running the same vague prompt through two different tools and comparing the near-identical results makes the averaging effect easy to see firsthand.

Common but wrong

NN/g's evaluation included a telling misfire. Asked for a profile page with a broad prompt, several tools produced social-media-style layouts; Replit's version emphasized email, contact details, and role, the pattern familiar from platforms like LinkedIn, while burying the content the page actually existed to deliver [6]. For a private course-attendee page, the most common pattern was the wrong one.

The trap is borrowed authority. Fluent, confident output suggests the tool knows best practice, but the default reflects frequency, not fitness. Best practice is contextual; frequency is not.

The working stance treats every AI default as one candidate among several: accepted deliberately when it fits, overridden deliberately when it doesn't, and never mistaken for a recommendation.

Polish is not quality

AI output arrives looking finished: real interface, real interactions, coherent styling. For two decades, visual polish was a costly signal of care, so it still reads as quality. AI decoupled the signal from the substance. NN/g's guidance is blunt: "don't confuse visual fidelity with design quality. A visually polished AI-generated mockup that aligns with your design system can still be a poor design" [7].

The evidence backs the warning. Even detailed prompts produced problems with hierarchy, grouping, contrast, and spacing, and the researchers remind builders that "AI prototypes are still just prototypes" [8]. A screen can be pixel-perfect and still answer the wrong question for its audience.

The practical shift is in the evaluation question. Instead of asking whether the output looks professional, the builder asks whether it serves the goal and audience it was made for. Polish answers the first question automatically; only judgment answers the second.

Borrowing taste honestly

Not every builder can articulate a visual direction from scratch, and there's an honest way to borrow one. NN/g recommends pointing to "established design styles or frameworks instead of using generic visual descriptions like simple, clean, and modern" [9]. Vague aesthetic adjectives feel like direction, but they leave the tool exactly where it started, reproducing the generic average. Moodboards, named styles, and admired products all give the model something concrete to execute.

The line falls at cloning. The same guidance warns against asking AI to imitate a specific brand, because famous brands don't guarantee good design; a design that works for a billion-dollar company can fail in a different context with different goals, audiences, and constraints.

A reference works as a source of decisions to notice, not a target to copy. Borrowing taste honestly means naming what a reference does well and asking for that quality, rather than asking for the reference itself.

Pro Tip! Moodboards and named styles transfer taste to the model; adjectives like modern and clean transfer almost nothing.

When common is correct

The strongest objection to this lesson is partly true: common solutions often are the right solutions. External consistency is a bedrock usability principle, because interfaces feel easier when they match what people have already seen elsewhere. NN/g concedes that AI's bias toward mainstream conventions "could potentially be good for the user experience" [10].

The same study's answer is that the tools often took this too far, producing indistinct, unpolished sameness. Familiarity earns its keep in some decisions and squanders it in others.

The mature version splits the decision space. Defaults are a fine floor for conventions, such as navigation placement and form patterns, and a poor ceiling for identity and fit, such as hierarchy, tone, and what matters most on a screen. Directing taste doesn't mean overriding everything; the skill is knowing which decisions deserve convention and which deserve a deliberate choice.

GenUI vs. vibe coding

Kate Moran of NN/g draws the field's cleanest boundary between two things that look identical from the outside. In generative UI (GenUI), the AI system decides to produce a visual or interactive element; in vibe coding, the person describes what they want and the AI builds it [11]. The difference is who makes the decision to build.

That decision determines what the AI is accountable for. Vibe coding holds the AI accountable for execution fidelity, whether the artifact matches intent, while GenUI holds it accountable for design judgment, whether the right element was generated at all. For a vibe-coded artifact, high quality means matching user intent while being usable, which is exactly why the person prompting needs a clear intent to match.

The implication closes the lesson. In vibe coding, the human initiated the build, so the design judgment seat belongs to the human whether anyone sits in it or not. When no one directs, the design decisions still get made, silently, by the averages.