The AI Gave the Website a 6/10. I Asked Who Taught It Taste.

I was looking at Luovi’s website because I wanted to understand the company better. Luovi is behind Helsinki Design Week and a lot of other cultural projects in Finland. The site is very quiet. Big images, a lot of white space, project names, not that much explaining. It feels more like an archive than a company trying to sell you something.

I gave the website to ChatGPT and asked it to review it. The answer was pretty confident. The projects were much stronger than the website made them look. The homepage needed clearer positioning. “References” should become “Work.” The project pages should explain what Luovi actually did, what the scale was and what the results were. There should be a proper team page. Better calls to action. A clearer way for someone to understand what Luovi could do for them.

It also gave the site a score. Six out of ten.

At first I mostly agreed. There were some obvious things that really were wrong. Typos. Old privacy text. A forgotten Squarespace template page. Those were easy.

Then I started thinking about the rest of the advice. Why exactly did the homepage need to explain itself faster? Why did every project need to become a case study? Why did a visitor need to understand immediately how to hire Luovi? Why should the site convert better?

I hadn’t told the AI that Luovi wanted more clients. I hadn’t told it the website was meant to generate leads. I hadn’t told it what Luovi thought the website was for. I had just asked it to review the website.

The AI had filled in the rest.

So I asked where those opinions came from. Some of the answer was simple. A typo is a fact. An old privacy policy is a fact. A page that still contains template content is a fact. But saying the site needs clearer positioning is something else. That is judgment.

The AI explained that this kind of judgment comes from all the patterns it has learned around UX, information architecture, marketing, design criticism, accessibility, agency websites and whatever else is represented in its training. It couldn’t tell me where one particular opinion came from. There is no source attached to “this homepage should explain the value proposition more clearly.” The opinion just appears.

That made the review much more interesting to me.

When we talk about AI training data, we usually talk about things. Which books did it read? Which websites? Which images? Which artists? But training data also contains judgments. This website is confusing. This one is clear. This design feels trustworthy. This navigation is bad. This case study needs numbers. This button should be easier to find. This is professional. This is dated. This is good.

The AI is learning those things too.

And then I give it a Finnish cultural production company and it tells me the website should convert better.

Once I noticed that, the whole review started to look different. The AI wanted the site to be clearer, more explicit, more measurable, easier to navigate and easier to act on. None of these are weird suggestions. They are basically the values that a lot of digital design has spent years teaching us.

Explain yourself. Reduce friction. Show evidence. Make the next step obvious. Don’t make people work unnecessarily.

Maybe that is actually good advice for Luovi. I still think parts of the site should be clearer. But Luovi is also not some SaaS company trying to turn anonymous traffic into signups. It exists inside a cultural world where maybe everyone important already knows Helsinki Design Week. Maybe not explaining everything is a kind of confidence. Maybe a very polished sales funnel would actually make the company feel less credible. Maybe the website is not trying to optimize for the random person who arrives from Google.

Or maybe the website is just old.

The AI doesn’t know.

I went back and changed the frame. What if Luovi doesn’t want to convert more visitors? What if the understatement is intentional? What if the audience already knows who they are?

The judgment changed immediately. Now the sparse homepage could signal confidence. The lack of aggressive sales language could fit the cultural positioning. The ambiguity could be part of the identity rather than a failure to communicate.

Same website. Same AI. Different idea of what good means.

That was the part I found interesting.

There wasn’t really one AI opinion about Luovi. There were a lot of possible opinions, and my question had selected one. When I asked for a website review without giving it a value system, the model supplied a default one.

Humans obviously do this too. If I show the same website to different designers, they will bring their own history with them. Where they studied, who they worked for, what kind of work they admire, which trends they hate, what they have seen succeed and fail. A Finnish cultural designer and a Silicon Valley growth designer could look at exactly the same homepage and see completely different problems.

The difference is that a human judgment can sometimes be traced back to a life. Someone can tell you they think something because they saw it fail on a previous project. Their professor taught them a particular way of thinking. They worked in an organization where everyone already knew the context. They have been in the room. Their taste has a biography.

AI taste has training.

That gives it a strange quality. It can draw on more examples than any one person could ever see, but it has never actually been responsible for one of those examples. It has never launched a website and watched nobody use it. It has never sat in a meeting where the rational design decision was obviously wrong for reasons nobody could really explain. It has never tried too hard to make something cool and then looked at it five years later and felt embarrassed. It has never walked around Helsinki Design Week and understood what kind of people Luovi actually wants to speak to.

It knows the patterns without having been there.

And increasingly we are using AI for exactly this kind of thing. Not just to make things, but to judge them. Review my portfolio. Which concept is better? What is wrong with this interface? Which photo should I use? Does this person look senior? Which idea should I pursue? Rank these candidates.

Judgment used to be expensive because another person had to spend time doing it. Now you can get endless judgment basically for free.

I wonder if that ends up having a bigger effect on design than the actual AI-generated images and websites. It would be easy to notice if every website suddenly had the same purple gradient and the same rounded cards. What is much harder to notice is millions of people asking similar models how to improve their work and getting answers that keep pulling things in roughly the same direction.

Clearer. Easier. More explicit. More proof. Less friction. Better hierarchy. Less ambiguity.

Every individual recommendation can be sensible and the overall effect can still be flattening.

The AI doesn’t need to design every website for websites to start becoming more similar. It only needs to become one of the things we ask whether the website is good.

I went back to Luovi after this and I still thought some of the original criticism was useful. The typos should be fixed. The old privacy information should be updated. I still think there are places where the company’s actual role in the projects is hard to understand.

But I didn’t really care anymore whether the site was a six out of ten.

The more interesting thing was how quickly a machine had looked at a website, produced a completely reasonable idea of what it should become, and how close I had been to accepting that idea without asking where its idea of good came from.

Connected reading

• The Edits Are the Training Data
• The Model Has a Voice of Its Own
• The New Aesthetics Will Not Be Prompted
• AI Slop Is the Factory. Brainrot Is the Product.

AI authorship note

This essay began with Kris Haamer asking ChatGPT to review Luovi’s website, then noticing that the model had quietly assumed the site should explain itself more clearly and convert visitors more effectively. Kris pushed the conversation from website critique into a question about AI judgment, taste, training, and the invisible value systems inside apparently reasonable design advice. AI supported the website review, development of the argument, drafting, and editing. Kris shaped the narrative, voice, paragraph structure, title, and final decision to publish.