Designing AI Features People Actually Trust

The first AI feature I shipped was technically impressive and almost unused. It auto-generated responses so fast that people didn't believe them — and they were right not to. Speed without legibility reads as magic, and magic is hard to trust with real work.

Since then I've settled on three rules. First, stream everything: a token arriving slowly and visibly feels honest, while a paragraph appearing instantly feels like a gamble. Second, make every AI action reversible — if the system edits a file, it shows a diff and waits. Third, show the system's reasoning surface: sources, confidence, what it chose not to do.

The uncomfortable part is that trust is a product decision, not a model decision. A smaller model that cites its sources will beat a bigger one that guesses confidently, every single time, in the hands of a real user.

In ReCode IDE we learned this the hard way. Early builds applied completions silently and beta testers disabled the feature within a day. The rebuild made the AI propose and the human apply. Adoption went from optional to habitual — same model, different contract.

AI is not a feature you add; it's a colleague you introduce. And nobody trusts a colleague who edits their work without asking.