Back like a decade ago when working on a design problem, I dreamt of AI tools for quickly exploring high quality design variants. At the time it wasn't there, but with the fast-paced recent advances in AI, it's basiscally here now.
The problem: Too much cancellation...
The trigger for this experiment was a confirmation dialog from a ride-hailing app. The original image example is from a LinkedIn post by someone, but it got lost in my feed, so I can't put credit where credit is due. Thus I have recreated the image just to make my point.
The dialog asks you to confirm whether you want to cancel your taxi ride – reasonable enough. But look at the buttons:
One says "Yes, cancel" – fine. The other says "Cancel" – but cancel what, exactly? The ride, or this dialog? A classic UX pitfall: using "Cancel" to mean both the primary action and the modal dismissal. Users in a hurry will tap the wrong one.
How to cancel... the ambiguity: Using Google Stitch to explore the solution space
Rather than sketching variants by hand, I fed the problem to Google Stitch by prompting "How would you improve the copy in this UI?" and asking it to generate four alternatives:
All four are improvements. None is perfect (if perfection exists...). That's exactly where human judgment comes in.
My version: taking the best from each
After reviewing the variants, I exported them from Stitch into Figma and designed my own version by cherry-picking what worked best across all four, adding my own human touches:
- Placing the modal at the bottom of the screen. To make it easier to reach on mobile, also with smaller hands.
- Icon which attracts attention, but doesn't signal error nor danger. Some of the icon suggestions of the AI is probably too aggressive with red warning triangles.
- Modal title is concise and aligns with the user action of potential cancellation. Vague, generic terms like "Confirm cancellation" or "Are you sure?" are harder to scan and understand without much thought. And "Keep ride" as the modal title would feel off, if you've just tapped a button to cancel the ride.
- Help text explaning a key consequence of cancellation – as proposed by the AI – instead of just restating what the title and buttons already communicate.
- Button labels are concise, consistent, and say exactly what happens. "Cancel ride" and "Keep ride" – no ambiguity, no cognitive load.
- "Cancel ride" button comes first – makes sense as flow, if the user indeed intends to cancel the ride. Also, if we had the Cancel button last (as proposed by the AI), users in a hurry might mistake thate for just dismissing the modal.
To mitigate the risk of unintended cancellations, the "Keep ride" buttons has a primary treatment (as the original design and proposed by AI).
We could have usability tested all four AI-generated variants to find the best solution – but applying human design judgment upfront lets us rule out the weaker options first, significantly reducing the test effort and saving precious time.
Btw, when a user cancels a ride, it would be an opportune moment to present a short survey asking why the user is cancelling – to gain valuable business and UX insights.
The bigger picture: AI + human designer = superpowers
What struck me most was was the speed of exploration. In the time it used to take me to sketch two alternatives, I had four fully-rendered variants to react to.
AI tools like Stitch don't replace design judgment. They compress the distance between "I see a problem" and "here are some directions worth reacting to." The hard part – knowing why one approach is better – that's still ours to figure out.
Ten-years-ago me would have loved this.
As a digital product organization you should definitely use AI to improve the UX of your products – but you still need the judgment and contributions from human product and UX designers. Which reminds me... I'm open to work ;)