The risks of letting AI dictate your design
I’ve been trying out a lot of vibe coding and vibe designing tools recently. If you’re not familiar with tools such as Lovable, Stitch and Magic Patterns they allow anyone to create designs, prototypes, or even working products in minutes using a simple chat interface. You tell the tool what you want and using the magic of AI it creates a design for you (in the case of Lovable, a design with working code under the hood!). I’ll walk you through a simple example using Stitch.
Let’s use one of the example prompts provided by Stitch: “ Make me an app for people who love skiing in the Alps “.
After confirming some pages to be designed, such as a welcome page, user registration page and resort selection page, Stitch waves its magic wand and comes back with a design for the app.
It’s returned a welcome page, a resort selection page, a (ski) run tracker page and a (ski) run details page. Pretty impressive stuff.
Look a little closer and you’ll see that the app is focused on tracking ski runs. The AI has interpreted, “Make me an app for people who love skiing in the Alps” as make an app for people who want to (in its words), “Record your ski runs, view stats, and share with friends”. Is this a need that people who love skiing in the Alps have? I’m not sure, but AI certainly seems to think so.
Let’s say that what you actually had in mind was an app to help people to browse the ski runs and facilities available at various resorts in the Alps. No problem, you can chat with the AI tool and simply request some changes to the design. For example, by telling it: “ Rather than recording and tracking ski runs, the focus of the app should be browsing facilities and ski runs available at resorts in the Alps. Update the designs to reflect this. “. AI once again waves its magic wand and some updated designs are generated.
This time Stitch has returned a better home page, letting users know that they can discover facilities and ski runs at resorts across the Alps. However, the run tracker page remains even though it was told that the focus of the app is browsing facilities and ski runs (you have to be very specific with these tools).
This repetitive loop of requesting design updates, reviewing the changes, requesting further updates continues until the AI has generated a design you’re happy with, or you’ve simply given up the will to live. These tools are called vibe coding, or vibe designing tools because you tell the tool what you want, but not necessarily all the details. You give it the vibe, and it does the thinking for you.
Yes, it can be frustrating having to endlessly request changes to a design, and yes a chat interface is a terrible way to describe design changes, but in theory these tools should enable anyone to quickly turn their ideas into designs, even working products. The problem is that not only is the workflow purgatory, it turns out that AI is not very good at turning prompts into well designed products.
What’s wrong with AI dictating your design?
If you ask a designer to create an app for people who love skiing in the Alps, they will come back with a hundred and one questions (see Why every design should start with the problem). What sort of people? What should the app do? What user problem(s) are you trying to solve? How sure are you that these are genuine user problems?
An AI tool on the other hand will simply make a lot of assumptions and diligently come back with a generic mobile app or website that loosely fits the bill. I guess that vibe coding loses some of its appeal if a user has to provide too much information upfront. The AI tool will always come back with a solution, but that solution will be far from optimal (hence the endless design updates) and as we saw with the ski run tracker example, might not even solve a genuine user problem.
Furthermore, because AI tools are trained on the most common mobile and web design patterns, they will always try to apply these design patterns. Every AI generated design ends up looking like a generic mobile app or website because these are the designs that AI knows best.
To effectively use these vibe coding and vibe designing tools you have to either be incredibly prescriptive, which kind of defeats the purpose of letting AI do the hard work, or you have to accept that getting to a half decent design is going to take a significant amount of time, and a lot of trial and error.
AI tools have their place, but I’m still far from convinced by vibe coding or vibe designing tools. Rather than an obstinate design master trying to dictate your design, I believe that AI should be an eager design apprentice, assisting with your design, but not trying to make key design decisions on your behalf (and certainly not bypassing, or even trying to replace designers).
How best to utilise AI for design
Design is certainly best left to designers, but that’s not to say that designers shouldn’t be using AI. AI is brilliant for speeding up the design process, and for assisting with design tasks and activities. For example:
- Using AI to help turn wireframes into UI designs and prototypes.
- Using AI to create and update design systems.
- Using AI to help analyse user research outputs, such as user interviews.
- Using AI to help create working prototypes from designs.
The use cases for utilising AI for design are endless, but one thing they all have in common is that the human is dictating the design, not AI. Using vibe coding and vibe designing tools to automatically generate designs might seem like a brilliant shortcut, but in reality they offer a never ending road filled with misinterpretations, generic outputs and poorly designed (and coded) products.
See also
- AI — designers’ friend or foe?
- Why every design should start with the problem
- Sorry, but we are not all designers
Image credits
Cyborg UX designer, created using ChatGPT
Originally published at https://www.uxforthemasses.com on June 16, 2025.
