Judging the output
The result is in front of the user and they have to decide whether to rely on it — usually without the time or expertise to verify it fully.
Stated confidence does not calibrate trust; observed accuracy does. Provenance and citations only help if they are cheap to check, and they now carry disclosure obligations as well as design ones.
16 patterns for this moment
Each one is a different way to resolve the tradeoff above.
Numbered or hovering markers that bind each claim to a retrievable source.
A visible estimate of certainty — level, band, or hedged language — so users know when to check.
An expandable record of steps, tool calls and intermediate findings behind the answer.
State when sources were updated and how much of the corpus was actually searched.
The system names exactly what a human should double-check, turning blanket doubt into one task.
A plain statement of a known limitation, placed where it changes the user's next action.
A closing summary of what changed, what was skipped, and what still needs a human.
A compact badge marking content as generated or assisted, at the point of consumption.
An expandable manifest: who made this, with which tools, and what changed since capture.
Layered disclosure on generated media: visible mark, embedded watermark, signed metadata.
A per-change record of which edits came from a human and which from a model.
Distinct visual identity for each acting agent, separate from the human who deployed it.
The agent chooses from a typed library of components and populates them; it cannot invent primitives.
When the agent needs input, it renders the specific fields it needs instead of asking in prose.
The agent assembles a layout of charts and panels answering the question just asked.
Substantial output moves out of the transcript into a durable side surface that can be edited and versioned.
Audit checklist for this moment
Walk one flow through these questions. Ticks are saved in this browser.