AI

AI interfaces answer back, and that breaks many of the assumptions classic UX was built on. A button does the same thing every time; a model may not. People have to learn what to ask, judge whether an answer is right, and decide how much to rely on it, often without any signal from the interface about how confident it is.

What separates good AI interfaces

Good AI products make the system's state visible. They stream responses so people see work happening, show sources so claims can be checked, and make it easy to stop, edit, retry, or undo. They set expectations early about what the tool can and can't do, and they keep the human in control of anything with consequences — sending a message, changing a file, spending money. The patterns are collected in the AI interaction hub, and many of them are old rules of feedback and system status applied to a new kind of system.

Where AI UX usually breaks

The blank prompt box is the most common failure: an empty field with no examples asks people to guess what the product is for. The second is false confidence, where a wrong answer looks exactly like a right one. The third is AI bolted onto an existing product as a floating button, with no clear relationship to the task on screen. And the fourth is lost work — long conversations that can't be searched, branched, or turned into something reusable.

This hub collects TYPENORM's articles and teardowns on designing for AI. The catalog lets you compare how leading assistants handle the same problems: ChatGPT, Claude, Gemini, and Perplexity, whose answer pages lead with their sources rather than hiding them.