Article
Notes on Designing AI Products
What changes when the product's core behavior is probabilistic: designing for uncertainty, latency, and trust in AI features.
Most product design assumes deterministic software: the same input produces the same output, and the designer’s job is to make that output legible. AI features break the assumption. The same prompt produces different answers, the model is confident when it’s wrong, and latency is measured in heartbeats, not milliseconds.
Working on AI products in Tokyo — including a 3D editor that generates scenes from natural language — has changed how I design three things.
Uncertainty needs an interface. When the system is 80% sure, the UI should look 80% sure. That means confidence-shaped affordances: showing alternatives instead of a single answer, making regeneration cheap and obvious, and never dressing a guess up as a fact.
Latency is a design material. A four-second generation feels broken with a spinner and magical with staged progress that shows the system working. Streaming, skeleton states, and honest progress copy are not polish — they are the product.
Trust is earned in the failure states. Users forgive a wrong answer they can correct. They don’t forgive a wrong answer they can’t detect. The design work that matters most is the correction loop: undo, edit-the-input, and show-your-work.
The tools will keep changing. These three won’t.