LLMs will always have a bias towards a specific style unless the model provider is injecting the entropy themselves and due to how human taste works that style will quickly become out of favor because it will be associated with lazy work.

They can be solved now with a follow-up prompt to remove the bad things, or preemptive rules to not do the bad things in the first place. That's different from "AI can't do hands" which was an intrinsic issue with the model and its robustness (there were tricks to mitigate it such as negative prompting, but far less comparatively successful than the web design analogy).

The alignment thing is just plain bad but a lot of these are like the em dash issue. It's simply overused.

It's super easy to generate content with AI and everybody's doing it but most content frankly doesn't really deserve to be generated or at the very least shown to the masses. There's nothing wrong with em-dashes but when people associate it with spam it directly devalues them.

Same thing with the badges, emojis, gradients etc. Your brain will simply start associating these tells as an indicator of cheapness just because the large majority of websites using it are just someone's side project that you probably shouldn't waste too much time engaging with.

This is a separate thing. It's not about imperfections: most of these websites work just fine. It's that everyone gets the same LLM, and if you just ask "hey LLM, make me a website", you get a very similar product. Eventually, people learn to recognize and hate that aesthetic because it's overused.

The exact same thing happens with LLM writing. There's nothing wrong with "it's not X, it's Y", but I not do my best to avoid that phrasing because it's tainted by AI slop. I did use it in this reply and part of me is bracing for the accusation...

You can overcome that by exercising more control, but most people don't know they can, and if you start micromanaging an agent, it slows you down and negates much of the speed advantage.

All the alignment issues seem likely to be solved as soon as LLMs can actually do vision well IMO.

Part of it is also lack of care, taste or vision from the slop-pilot. Sure the clanker designed and built the UI, but someone looked at it and went "Hell yeah! Ship it!".

Developing good UI/UX is hard, like really hard. Very few people are good at it and if you're not good at it, what makes you think you're able to recognize a good UI when you see it? So much thought and and foresight goes into a good UI that it requires a serious amount of prompting. When I've worked with actual UX/UI designers, my job has been that of the LLM. The designer spends ages on planning and understanding, then prompts the developer to create the actual UI and iterated endlessly.

Industry most hasn't cared about UI/UX since the 1990s and even if it started to cared now, employment is fairly safe, because it was always mostly research, labs and prompting. It's just that now we can iterate faster because prompting the LLM is faster than prompting your five .Net developers hidden away on the second floor, in the office behind customer service.

I think it's from laziness and not having skill at design. Everything is fixable already,ai cant read your mind.

AI use seems to mostly show what people care about. If something matters, you'll work carefully on it. If it doesn't have chatgpt do it.