AI products need a different UI rulebook, not the old one with a chatbot bolted on.

There is a framework going around with 39 principles for designing human-AI interaction. The starting premise is the part worth sitting with. AI is probabilistic, not deterministic, so trust, transparency, and recoverability matter as much as raw model capability.

Most of the principles are things good interface design already knew. Show your evidence instead of a confidence score nobody can interpret. Keep outputs editable. Design for graceful failure instead of pretending failure won't happen. Limit autonomy based on actual risk, not on how impressive the demo looks.

What's different is the stakes of skipping them. A deterministic interface that hides its reasoning is just unclear. A probabilistic one that hides its reasoning teaches the user to either blindly trust it or ignore it entirely, and both outcomes are worse than a bad button label ever was.

Which of your product's AI features would fail a basic "can the user tell why it did that" test right now?