Why AI-generated UI is average

There are numerous posts on how fast you can generate interfaces now. But very few of them talk about the visual design side of things, arguably the most visible artifact of the whole process. That part is underrepresented, or worse, misrepresented.

And it makes sense why. Visual design is often treated as a nice-to-have, a layer of polish you apply at the end rather than a discipline with real business value. So when AI starts generating interfaces at speed, the unassuming eye is impressed. And what happens is predictable: it gets over-hyped by those who aren't fluent in visual design, and quietly ignored by those who are. (because everyone is focused on shipping products now, right?)

So… what's visual design?

Coming from an agency and freelancing background, I've spent the better part of my days designing, directing, and evaluating the quality of the interfaces we create. And if there's one things that's true about visual design, it's this:

it is the summation of all the minor and major visual decisions you make, from the smallest icon choices, the way your button animates on hover, the way you speak to the customer, all the way to the structural layout within any given medium, with a goal to elicit a certain emotion or drive a business outcome.

And the part that many people neglect, is the duality of visual design as a practice:

Visual design is composed of both explicit and tacit knowledge. Many designers, and AI models alike, are only decent at the former, and often ignore the latter. And that's the recipe for a sea of average-looking interfaces.

To describe what explicit knowledge is; something becomes explicit when it's easy to document, share, and transfer within an org. Think about how many variations of a button do you need? What viewports to design for? How do you organize your variables library?... Simply put, it's the equivalent of agreeing to what "the sky is blue" is for your organization.

Tacit knowledge is quite a diferent story; It is the knowledge that books won't give you, nor is it knowledge that comes from an instruction manual. Mostly earned from having worked with ambiguity for a while, and typically the result is the development of a unique point of view. A perspective.

A perspective in visual design becomes highly important, because although you might be targetting the same customer as the brand nextdoor, if you heavily rely on explicit knowledge, you will end up with a what I like to refer to as, a slopsite.

If everything starts to look the same, your main differentiator becomes your heritage and reputation, and in a competitive flashy market your chance to stand out and gain trust is close to zero. You have rendered visual design as useless, and are metaphorically racing on one leg.

On the topic of taste

I think it would be a major disservice to the food industry to talk about 'taste' without referencing the chef.

Similar to the works of a chef, everyone mostly has access to similar ingredients (the small and big decisions). But a chef, through rigorous iteration and understanding his ingredients very well, knows the proper combinations that elicit the exact flavour, taste, and presenation of a dish. Combined, these help shape the consumer's experience. (mostly emotional)

That's similar to how it works in visual design, too. There's a detail worth noting, however. The same combination of design ingredients, may evoke completely different set of emotions in different parts of the world. And that part, is mostly tacit.

Why is this a problem?

Back to the AI-gen visual design. Why am I attempting to be specific about what visual design is as a practice?

It's because I keep seeing many designers (and non-designers alike) use AI as a police sketch artist.

They try to describe how they want their website to feel, what feeling they want to evoke, like "premium," "a hug" or "minimalist..." completely forgetting that this is not what a language model is made for.

This is the way you would chat with a peer who has a shared context, a client, a trusted family relative... the common denominator here is that these are all humans with feelings. They will likely understand what feeling you're describing and be on same page with you, unlike a language model.

A language model, on the other hand, averages out that feeling with the vast data that it was trained on.

Even if it had good reference points on what "minimalist" feels like visually, it will likely resort to middle-of-the-bell-curve slop, which generates visual design that is average at best. something that the majority of the viewers would be 'ok' with.

I was chatting with a friend of mine in the AI space, Michael Zakhary (who also happens to publish on Medium on Machine Learning) and he confirmed this occurence and introduced me to this phenomenon called "Modal Collapse."

I won't go into detail on what Modal Collapse is, that's what google is for, but here's a simple alanogy: Imagine asking a student to draw random animals. Instead of drawing dogs, cats, birds, etc., they just draw the same dog over and over because it always gets a passing grade.

And this is exactly what happens with AI in visual design. The passing grade is the average person's taste (not the designer's), and we know that being a designer is not the most common on this planet, and what ends up happening is you get middle-of-the-bell-curve slop. And from a designer's POV, this just sucks.

"Can you please make it look human?"

The way we have humanized AI, gave it a name, a chat interface, and a speech pattern that's just. so. re-affirming. That is the source of the trap.

AI doesn't understand you. AI doesn't really feel what you're feeling. And treating it like it does, is simply foolish.

You can ask AI to build something so premium, on-brand, minimalist, large type, and playful color palette but still get a generic outcome (around 90% of the time) because you're leaning on feeling-heavy words. In other words, what you're saying is open for interpretation.

However, if you try commanding it to do something, such as to make this button round, use a light border outline on the input fields, or organize a design system for you... Suddenly, AI has become your best ally. And it's only because you're commanding it to do something explicit. And because that's the right way to ask.

Designers, leaving the driver's seat is not the brightest idea.