AI interfaces need a point of view

The default AI screen is a transcript. A box, a stream of words, a cursor that blinks as if motion were the same as progress. It is a reasonable prototype. It is a poor product. A transcript hides the thing a person most needs: a point of view about what the software believes, what it does not know, and what the person is still responsible for.

A transcript hides the job

Chat is a good input. It is rarely a good record of work. People do not want a relationship with a panel. They want to finish a review, compare two options, or leave with a draft they can defend. When the interface only shows the conversation, the job stays trapped inside it.

A stronger screen names the artifact. The summary, the checklist, the diff, the decision. The conversation can sit beside that artifact. It should not replace it.

States a person can trust

I keep a short list when I sketch these interfaces. If a state is missing, the screen is pretending.

• Waiting — the system has the request and has not started speaking.
• Streaming — words are arriving, and they are not done.
• Low confidence — the answer is partial and says so in the interface, not in a footnote the model might forget.
• Source missing — a claim that needed evidence does not have it.
• Person editing — the draft now belongs to the human, and the model does not keep typing over them.
• Done, with a way back — there is a result, and a clear path to revise it.

None of this requires a theatrical dashboard. It requires labels, contrast, and the discipline to leave a gap when the system should not bluff. Motion, if you use it, should explain arrival. It should not perform intelligence.

Put judgment where the work is

Engineering judgment shows up in boring decisions. Can the person edit a sentence without re-running the world? Can they see which part came from a source? Can they reject a suggestion without feeling they broke the spell? If the answer is no, the interface has taken a job it cannot hold.

Confidence is a visual choice. If everything looks equally sure, the product is lying.

A critique you can run

Before I call an AI screen finished, even as a prototype, I walk this list. It is the same list filed in the lab as a learning note.

• What is the artifact, other than the chat?
• Which state is visible when the model should hesitate?
• What can the person change with their hands?
• What happens to a claim with no source?
• What did we refuse to automate?