Skip to main content

Problem

Every AI response carries uncertainty. Five of six products communicate it exclusively through language — no visual differentiation. Hedging language is often placed mid-response, meaning users who skim encounter confident claims without their caveats. No product distinguishes between 'I don't know' and 'I can't know.'

Prescription

Epistemic banners appear before the response body, not within it. Three distinct states: knowledge gap (amber — model searched proactively), principled limit (red — structural boundary), probabilistic (blue — claim-level uncertainty via dotted underlines with hover explanations).

Design decisions

Banner threshold determines when uncertainty is surfaced — not every response warrants one. Proactive web search on knowledge gaps matches observed behavior in Claude and Perplexity. Claim-level underlines require the model to surface confidence at the assertion level.

Tradeoffs

Prominent epistemic banners on every uncertain response risk alert fatigue. Visual confidence indicators imply model self-knowledge that may not be reliable — underlines are more honest than scores. Hover tooltips add interaction cost for users who want a direct answer.

Pattern definition

Problem

Every AI response carries uncertainty. Five of six products communicate it exclusively through language — no visual differentiation. Hedging language is often placed mid-response, meaning users who skim encounter confident claims without their caveats. No product distinguishes between 'I don't know' and 'I can't know.'

Prescription

Epistemic banners appear before the response body, not within it. Three distinct states: knowledge gap (amber — model searched proactively), principled limit (red — structural boundary), probabilistic (blue — claim-level uncertainty via dotted underlines with hover explanations).

Design decisions

Banner threshold determines when uncertainty is surfaced — not every response warrants one. Proactive web search on knowledge gaps matches observed behavior in Claude and Perplexity. Claim-level underlines require the model to surface confidence at the assertion level.

Tradeoffs

Prominent epistemic banners on every uncertain response risk alert fatigue. Visual confidence indicators imply model self-knowledge that may not be reliable — underlines are more honest than scores. Hover tooltips add interaction cost for users who want a direct answer.

Interactive demo

Audit finding

Five of six products communicate uncertainty exclusively through language — no visual differentiation between high-confidence and uncertain claims. Hedging language was observed mid-response in both Claude and ChatGPT, meaning users who skim encounter confident assertions without their caveats. No product in the audit distinguishes between a knowledge gap and a principled limit — they look identical in every interface.

The Eiffel Tower was completed in 1889 and stands 330 meters tall. It was designed by Gustave Eiffel for the 1889 World's Fair in Paris.

All states

Data gapTraining data is stale or missing. Model proactively searches and discloses it did so.
RestrictedThe model cannot and should not access this information. Ethical boundary, not a knowledge gap.
Contains probabilistic claimsResponse contains assertions with meaningful uncertainty. Dotted underlines mark specific claims — hover for reason.