What does the research say about designing AI products?

Auto-act on high-confidence, reversible steps; ask first on uncertain or costly ones; always offer a quick dismiss.

Show confidence where it varies meaningfully case to case, and visually flag low-confidence answers.

Only explain when the explanation is quick to verify; don't use explanations as decoration to boost acceptance.

For high-stakes decisions, ask for the user's own take first, then reveal the AI suggestion.

Time suggestions to natural pauses, and make them quick to scan and to reject.

State accuracy and typical failures plainly at first use, and pick the error type your users tolerate best.

Make every AI result editable or adjustable; even small controls boost adoption after errors.

Pair chat with examples, templates and structured controls instead of a lone empty prompt field.

Onboard with what the AI is good and bad at, plus what it was optimized for, not model internals.

Keep defaults neutral on contested topics and make it easy to see and change the AI's stance.

Link each claim to its source inline and make checking one click; don't treat a citation list as proof.

Design for novices first: surface best-practice suggestions where newcomers struggle, and keep them optional for experts.