Designing AI features people actually trust

A probabilistic system behind a deterministic-looking interface is a trap. The UI implies certainty the model does not have, so the first time it is confidently wrong, the user does not conclude 'the model made an error' — they conclude 'this product lies to me'. That judgement is very hard to reverse.

Give uncertainty a design

On a nutrition scanner I worked on, the model was strong but imperfect. Rather than hide that, we built three distinct result states: confident, uncertain and unrecognised — each with its own visual treatment and its own next action.

The uncertain state does not guess. It says what it thinks it sees and asks one clarifying question. That single pattern did more for trust than any accuracy improvement could have, because a product that admits doubt is a product you can believe when it does not.

• Confident: give the answer plainly, reasoning one tap away.
• Uncertain: state the doubt, ask one question, offer manual entry.
• Unrecognised: fail clearly and fast, with an obvious alternative path.

Make being wrong cheap

Every AI feature needs a two-tap correction path. Not a support form, not a thumbs-down that vanishes into a dashboard — an immediate, visible fix. It turns the product's worst moment into its best feedback loop, and it tells the user they are in control of the system rather than the other way around.

Stop starting with a chat box

Chat is the default AI interface because it is the easiest to build, not because it is usually the right one. An empty text field asks the user to figure out the product's capabilities unaided, which most people will not do. Structured inputs, suggested actions grouped by job, and inline assistance inside the existing workflow beat a blank prompt almost every time.