The chat box is an extraordinary interface. It accepts ambiguous input, lets intent arrive before structure, and gives a product room to respond in kind.
It is also where many AI products stop designing.
Language is input, not the whole interface
The moment an AI system can change state, spend money, contact someone, alter a file, or carry memory across sessions, a conversation alone is rarely enough.
People need to know:
- what the system understood;
- what it plans to do;
- what information it used;
- which parts are uncertain;
- what has already changed;
- and how to interrupt, correct, or reverse it.
These are interface questions. Some answers can be expressed in language. Many are better as previews, selections, timelines, permissions, diff views, or persistent state.
Confidence is not trust
During experiments with an instruct-model Android launcher, the gap between “I understood the request” and “I should execute this action” became very concrete. A model can sound certain while its plan rests on a wrong assumption.
Trust comes from legible behaviour: appropriate friction, visible boundaries, good defaults, and recovery that does not punish the user.
Design the seam
The most interesting AI interaction work is at the seam between probabilistic interpretation and deterministic action. That seam needs an interface.
The future is not chat versus GUI. It is products that know when language is the best material — and when a person deserves something they can see, compare, manipulate, and control.