Designing for Trust: UI Patterns for AI Products
A capable model behind a confusing interface still fails the user. Some of the most important design work on AI products happens around the model, not inside it.
Showing sources and citations, even briefly, does more for trust than almost any other single change we've made. Users don't need to read every citation — knowing they exist changes how much they're willing to rely on an answer.
Streaming responses matter for more than perceived speed. Watching an answer form gives users a natural point to interrupt or redirect if the response is heading somewhere unhelpful, rather than waiting for a complete — and potentially wrong — answer.
We also design explicit uncertainty into the interface: a visibly different state for 'confident answer' versus 'best guess, please verify.' Pretending a system is more certain than it is erodes trust faster than admitting the limits upfront.
