Don't fire your designer: UI, UX and LLMs

Designers have it rough. My wife Martha and I have been working on and off together for over two decades as a designer / programmer pair and in the past, one of the things that stood out is that while clients would be fairly deferential to me on development decisions, there was very little of the same automatic trust when it came to design decisions. Everyone seems to have strong opinions on design.

If I asked a client microservices or monolith, very very few would have the expertise to make a better informed decision than I as a consultant, and very few would even deign to have an opinion (unless they read too many blogs, in which case microservices is the only architecture). When it came time to pick a font, color scheme, or layout however, you can bet the client had an opinion, and often quite justifiably so. Taste is personal and very few clients would choose something that doesn't match their taste even if a designer says it'll work better. Unfortunately design also benefits from simplicity, and sometimes there are too many stakeholders in a room to end up with a good, simple design (case in point - the number of carousel-homepages that exist - a decidedly "we can't make a decision on what's important" outcome)

What's interesting however is that as many of the simpler aspects of the design profession are being commoditized by AI - from event posters to my-nephew-could-have-done-it websites and logos - the stark sameness of all of these designs is quite apparent.

My unscientific guess is this sameness comes from the fact that good design is both exceptional (it looks unique) and well-structured (it follows rules) and LLMs, which are sort of like an average-face-morph of the designs littering the internet, can't possibly keep all those exceptional designs in their weights, so really all you get is the well-structured part, and sometimes not even that, as the average of two great type-hierarchies isn't a great hierarchy. The bit of design work from the RLHF these models go through isn't being done by great designers and there's likely a limited amount of great design that's fed positively back through the system.

What's funny is these designs often appear decent and unique the first time you see them - "wow - I got Claude to design something really special" - but then like the Battlestar Galactica Cylons, there's a limited number of them and so within each type, the designs all look the same.

Some of the very clear tells that I've seen include: serif followed by a different color sans-serif title (or vice versa). That one beige color. Lists leading with 0 (e.g. 01, 02, 03). The list goes on and on, and as someone with the time to do a deeper analysis summarized:

The loudest complaint is not any single feature. It is that the sites are recognizable on sight. "They all look the same" and "screams AI / slop" each show up in about 13% of on-topic posts. (vibecoded-design-tells)

Of course this person has also created a Claude Code skill to remove the "slop" ... with AI. Having played around with it and looked at some examples, it does seem to work, but the result is just more genericity - even if it's not readily apparent that it's AI-sourced genericity. You're not getting that unique and beautiful unicorn.

Believe me, I've tried - I've done dozens of designs in Claude code and have had trouble getting something that doesn't fit the pattern out, even when giving it more and more detailed text specs. If you feed it images, it can copy just fine, but you're not going to get something unique and exceptional with a copy (by definition).

All that is to say that a good designer is still going to have a job and it's one that actually has become more valuable not less. As more people opt for "good enough" LLM-based generic designs, that tiny spark of human creativity is going to stand out in the crowd as special.

The current cykod.com site was designed by Martha (in about a day) and I think, while still looking current and modern, doesn't have that unfortunate patina of AI design. I wasted a few days churning on various different concepts to try to get a jump-start, but all of those resulted in different models of LLM-slop that didn't help the final product. Now, of course, it was actually 100% implemented and connected into the Headroom CMS by Claude. But we all know LLMs are great at coding, so that shouldn't surprise anyone.

Having made the case that LLMs aren't great at design - the question is how about the UX? Most apps' UX has always been mediocre, so seeing haphazard UX in the wild it's tough to know if it was the LLM or the human.

I have a little less of a definitive opinion, as I personally may just be unable to get LLMs to properly consider the UX of what they're building. Again, believe me I've tried. All the normal tricks - feedback loops, screenshot visual reviews, user stories, user roleplaying, clear-context review agents - all have effectively failed to create great multi-page UX that really considers the user. Halfway-passable UX is the best I've been able to accomplish for long, unmanned runs, and I've come to the conclusion that it's not worth it. The amount of debt that comes due all at once when trying to review full LLM-designed-and-built flows is brutal. This is even when providing screens and write-ups describing those screens.

Instead, I've switched to a screen-by-screen approach when building UIs with LLMs and have had significantly better results. More on that in another post, but suffice it to say - humans still need to be responsible for good UX in websites and apps, because LLMs aren't going to do it for you (yet at least). So, if your designer is good at UX, and good at design and UI they should have great job security for the foreseeable future. If you already fired them because of AI-psychosis, it might be time to go begging to get them back. Even if you haven't noticed, your clients and customers have. And it's likely not pretty.

P.S. If you think you have a wonderful, unique snowflake of an AI-designed site, I'd love to see it and hear about it, however just be warned there's probably more than a few of the same model out there ;)