AI Has a Design Problem

From the newsroom: Anthropic’s (Claude) Fable 5.1 and ChatGPT’s GPT-6 ‘Astra’ have been launched. Dun dun dah.

In the earlier stages of the AI adoption cycle, releases were highly anticipated and very much centre stage. Now, it seems that with every new model release, one of, if not the first, things mentioned is how the latest model(s) bring a host of changes to ‘fix’ issues and that they now rank higher with an ‘X’ score on the ‘Y’ benchmark. Throw in a “We’re entering a new era”, “AGI is here”, blah-blah statement, and that’s a wrap. It’s so predictable how these go. I think they might be using the same script because you can bet this gets followed up with: “It’s so powerful, we’ve had to limit ABC for security reasons.” I’m looking at you, Fable, and your ‘Mythos’ clan.

Speaking of Claude, have you had a chance to experiment with its design tool? This was recently deployed into Claude Code with the /design command.

At first, I was blown away, and while I still find it helpful, of late it does seem to have become a bit ‘samey’. The similarity between designs has drastically increased, and there’s a feeling like it’s shackled to ensure it ‘plays it safe’. Results will (naturally) vary by model; the prompt; whether you have a custom design system; image references; the weather; whether you have prayed to the AI gods for wisdom or decided to down a packet of sherbet to work at twice your normal pace. Missing any of these? Prepare for the same outputs.

It’s a double-edged sword. The implementation of the ‘design system’ functionality has made the output more reliable, though the trade-off is in how unique those outputs are. It’s in beta, so it is likely to change – let’s hope for the better.

Interestingly, in my attempts to ply the design system with imagery to vary the results, I stumbled upon a paid-for resource called ‘Mobbin’. I’ve found it to be quite handy. While some might sneer at it being paid-for, I will say that it’s not a dump of random design work; it contains ample screenshots from top-quality sites and apps, which are well catalogued.

Before you think of comparing it to Dribbble, Behance, Pinterest or the like, the purpose of Mobbin is to provide screenshots of the varying states of an application. In some cases, there are apps with 150+ screenshots walking you step-by-step through all the options. For me, this is seriously valuable and very enlightening. Want to give it a go? Use my link for 10% off the ‘Pro’ or ‘Team’ plan.

Armed with these resources, I hoped that by providing examples to Claude Design, it would help steer and influence the final result. While it does work to a certain degree, there seems to be a design bias which remains rather stubborn and sometimes challenging to override. Maybe it’s so you have to spend more tokens…

It’s about now that someone is thinking how “he’s not doing it right” and that I need to “check out XYZ on YouTube to get it”.

Let me be clear, I’m not perfect (despite my best efforts). AI is a huge part of my daily workflow and habits. I’ve seen lots of videos on YouTube of gurus giving “tips” and “prompt guides” which claim to help you deliver “awesome and unique designs”. Anyone else find it funny how most of these videos rely on expensive models like Fable?

That being said, if you can get something out of prompt-mashing sessions, you need to remember that just because AI can produce a “design”, it doesn’t necessarily mean it’s good. Do we really believe the AI engages in design thinking? Sure, it can arrange components into something visually plausible based on what its LLM has previously reaped, but I wouldn’t necessarily believe it understands the underlying UX or experience. That’s where your experience comes into play.

If I had to give reasons ‘for and against’ AI design:

• [Pro] Extremely fast for concepts and exploration
• [Pro] Lowers the barrier between an idea and something visual
• [Pro] Great at bridging a gap between concept, prototype and early build without causing too much friction to get started
• [Con] A good-looking design doesn’t guarantee it’s good for users
• [Con] Outputs are often very similar; there’s a design bias that makes it easy to spot AI-generated designs
• [Con] Iteration, consistency and limits can become frustrating - prepare to run out of tokens quickly…

For some, Claude’s design feature is great. If you are working on internal tools or a proof of concept, it’s a solid option. At the beginning stages of projects, having something to quickly visualise and demonstrate can be very valuable. Typically, that’s the stage where you are on a shoestring budget, and it could be the difference between being able to pitch for investment or showcase for awareness. It’s not a silver bullet.

What about you? Have you experienced a similar-looking design (or common design elements) across projects? Did you manage to overcome it, and if so, what did you do differently?