DeepProgrammer

LabLearning notePrototype

Notes on showing uncertainty

A learning note about the states an AI interface should show before anyone calls it a product.

6 Nov 2026

Status
Prototype
Kind
Learning note
Role
Designer-engineer taking notes. Not a client engagement.
Scope
A critique checklist and a set of interface states. No production traffic, no measured lift, no fake screenshot of a dashboard.

  1. 01 Waiting

    The request is held. Nothing pretends to be done.

  2. 02 Streaming

    Words are arriving, and the screen says they are unfinished.

  3. 03 Low confidence

    The doubt is a label, not a vibe.

  4. 04 Source missing

    A claim without evidence is marked, not polished.

  5. 05 Person editing

    The draft belongs to the human. The model stops typing.

  6. 06 Done, with a way back

    There is an artifact, and a clear path to revise it.

Problem

Concept demos of AI products tend to show the happy transcript. The useful design work is the moment the system should hesitate.

Approach

I listed the moments a person needs the screen to be specific: waiting, streaming, low confidence, a missing source, a human edit, and a finished artifact with a way back.

Stack

Notes · Interface states · A critique checklist

What changed

I stopped treating a streaming cursor as evidence that the experience is working. The notes now start from the artifact and the doubt around it.

What I learned

If every state looks equally sure, the interface is making a claim the model cannot keep. Labels do more than decoration.

Public learning notes. Not a case study, and not a description of any private product.

Related note: AI interfaces need a point of view · Build