AI Can Copy Good Design. It Cannot Recognize It.

There are two halves to this job. Knowing what to ask for, and knowing what you got back. Nearly every conversation about AI and websites is about the first half, because the first half can be written down. The second half decides whether your site works, and it is the one nobody can hand you.

It takes direction beautifully. From whom?

Every defense of an AI-built site runs through the same move. Sure, but you can direct it. Tell it to be warmer. Tell it the hierarchy is wrong. Move the button. All true, and it takes correction better than most people expect.

But direction has to come from somebody, and that somebody is usually a person who has never designed a website, never watched a stranger try to use one, and has no particular reason to know what separates a page that converts from a page that just sits there.

That is not an insult. There is no reason a person who runs a plumbing company or a law practice should know this, any more than I should know how to re-pipe a house. It is a specialty. The trouble is that AI makes the absence of it invisible, because it follows weak direction exactly as smoothly and exactly as confidently as good direction, and hands you something that looks finished either way.

A brief gets you halfway

Watch what people actually type into these tools and it is some version of: modern, clean, professional, trustworthy. Maybe a competitor's URL. Maybe a color they like.

Every one of those is an outcome, not an instruction. "Professional" is something a visitor feels. It is not a thing you can build. What builds it is fifty small decisions the brief never mentioned, which the tool then made on your behalf.

A real brief looks nothing like that. It says: the person landing here is stressed and comparing three options in one sitting, so the proof goes above the services rather than below them. The phone number is the primary action, not the form, because these people want a voice. Two of our six services make almost all the money, so the other four come off the homepage entirely.

Hand that to AI and you get a genuinely different site. Not a prettier one. A more effective one.

And you are still only halfway, because now something exists and somebody has to look at it.

The eye is the part that cannot be written down

This is the half that gets skipped, and it is the more important one.

Designers have an eye. That sounds like mysticism, so let me be specific about what it actually is, because it is not mysticism and it is not taste in the sense of having opinions about colors.

It is trained discrimination. It is the ability to look at something and know that it is wrong slightly before you can explain why. It gets built the boring way, over thousands of hours of making things, being told they are not right, looking at work that is better than yours, and slowly developing an internal alarm that goes off when something is a few pixels or one word or half a tone away from where it should be.

Some of what that alarm catches:

• Optical versus mathematical. A play triangle centered inside a circle, perfectly, by the numbers, looks like it is sitting too far left. It is not. Your eye reads visual weight, not bounding boxes. A designer nudges it a couple of pixels right without consciously deciding to. Nobody can articulate the rule, and everybody can see it once it is fixed
• Proximity, which is doing more work than anything else on your page. A heading belongs closer to the paragraph it introduces than to the paragraph above it. That single relationship is most of what makes a page feel organized. Generated layouts almost always space everything evenly, which quietly tells the reader that nothing belongs to anything, and the page feels flat for reasons nobody can name.
• Emphasis that has cancelled itself out. If four things on the page are bold, nothing is bold. This is obvious once said and almost never caught by the person who added the fourth one, because they added it for a good reason.
• A headline one word too long. It wraps onto a second line and the entire balance of the hero changes. A designer cuts the word. The tool has no reason to, because the sentence was fine.​
• Contrast that passes and still fails. Text can clear a contrast checker at 4.5 to 1 and still be miserable to read at fourteen pixels on a warm background. The checker says yes. The eye says no. The checker is not wrong, it is just answering a narrower question than the one that matters.

Here is the thing about every example above: you can see all of them once they are pointed out. That is exactly the problem. Recognition is easy. Detection is the hard part, and detection is what an eye is. Being able to agree that something looks better afterwards is not the same skill as being able to find it in the first place, and the gap between those two is most of a career.

A recipe is a brief. It gets you most of the way there. But any cook worth eating from tastes the thing before it leaves the kitchen and adjusts, and no recipe has ever contained the instruction that mattered at that moment.

AI can copy good design. It cannot recognize it.

This is the part I want to be precise about, because it is easy to say badly.

AI can absolutely produce good-looking work. It has seen enormous amounts of excellent design and it can reproduce the surface of it convincingly. Show it something and ask it to do something similar and it will, often impressively.

What it cannot do is discriminate. Ask a model whether a design is good and you get a fluent answer. Confident, well organized, correct vocabulary. Ask it the same question about the opposite design and you get an equally fluent answer defending that one. That is not judgment. That is fluency, and the two are extremely easy to confuse, particularly if you do not have the expertise to tell them apart, which is the entire situation this article is about.

It will also simply agree with you. Tell it the layout is weak and it will explain why the layout is weak. Tell it the same layout is strong and it will explain why it is strong. That is not an opinion changing. That is compliance, and compliance is the opposite of what you want from a designer. The value of a designer is substantially in the moments they tell you no.

Three things stay true here no matter how good these tools get, and they are worth understanding because they are structural rather than temporary:

1. It has no stake. It never finds out whether you got the customer. No outcome ever reaches it, so no outcome ever corrects it. It will make the same choice next time with the same confidence.
2. It has no context. It does not know that every competitor in your city went navy blue, or that your best clients arrive by referral already half-convinced, or that the question people actually ask you on the phone is the one thing missing from the site.
3. It is an average by construction. It produces the statistical middle of everything it has absorbed. Good design is, by definition, not the middle. That is what makes it good.

The things nobody knows to check

Past the visual, there is a whole category of things that never come up because it does not occur to anyone that they should. Whether the interior pages make sense to someone arriving from Google, since most of your traffic never sees the homepage. Whether form labels and alt text exist at all, which are among the most common accessibility failures on the web and in California can become a legal question as well as an ethical one - more on that in website accessibility lawsuits in California, since we build websites rather than practice law.

Whether strangers can read your customers' data, which is not a thought that occurs while you are trying to get the header to look right. Researchers at Escape scanned over 5,600 public apps built on vibe-coding platforms in 2025 and found more than 2,000 vulnerabilities and over 400 exposed secrets.

And whether the contact form email actually arrives. Almost funny, except it is the single most common thing I find. The form submits, the success message appears, the email lands in a spam folder nobody checks, and it has looked like it was working since March.

Why you cannot review your way out of it

The natural response to all of this is: fine, I will just look at what it made and fix what is wrong.

You cannot, and this one is not about skill at all.

You already know where everything is. You wrote the words or approved them. You know what the third menu item means and which button matters. Every one of those facts sits between you and the experience a visitor has, and you cannot put them down. The person who made the thing is structurally the worst available judge of whether it is obvious.

True of me too, which is why nobody serious signs off on their own work without putting it in front of someone who has never seen it. Nielsen and Landauer established in 1993 that five fresh users surface roughly 85 percent of a design's usability problems. The finding has held for thirty years, and the important part is not the number. It is that watching people is the mechanism, and there is no substitute route to it.

For a sense of how much hides in plain sight even with professionals involved: Baymard's benchmark of 50 leading e-commerce sites, companies with design teams and real budgets, found an average of 39 identified areas for improvement in the checkout alone. Thirty-nine, with a team. The question was never whether your site has problems.

The instrument is not the practice

So no, the answer is not "use AI, do what you like, you are done."

A better camera has never made anyone a photographer. What it does is let a good photographer work faster and let a bad photograph arrive sooner.

Here is what actually happens when someone who does this for a living uses these tools. They generate something and reject it. Adjust, generate, reject again. Somewhere around the sixth or seventh version something is close, and then they spend longer fixing three specific things about that version than they spent generating all seven.

Most of the output gets thrown away. That is not waste. That is the work. Producing options was never the valuable part, and AI now produces infinite options for nothing. The value is entirely in the rejecting, and rejecting requires an eye.

Someone without one generates a site, sees something that looks acceptable, and ships it. Not out of laziness. Because nothing told them to keep going, and nothing was ever going to.

Use the tools. I do, constantly, and I would not go back. Just know which half of the job you are doing with them, and get the other half from somewhere.

If you have something built and you are not sure what you are looking at, that is a specific and answerable question, and it is usually a much smaller job than starting over. It is what our AI site rescue is for. The wider position is on AI and your website.