The confidence gradient
Showing uncertainty in typography, not in hedging prose.
Confidence is uneven. A single answer can have parts the model is sure of and parts it's guessing at, woven into the same sentence. Most products render that sentence with uniform typography, so the user has no signal at all. The confidence gradient solves this with the smallest possible toolkit: weight, color, italics.
"Hedging doesn't have to be a sentence. It can be a type weight."
Three levels, minimum of color.
Break the answer into spans: sure, likely, unsure. Render sure spans as full-weight body text. Likely spans in a slightly dimmer color. Unsure spans in muted italic. The reader can feel the confidence gradient without having to parse a caveat.
Avoid the loud version of this: red/yellow/green highlights over every sentence. That turns an answer into a traffic warning. The cost is high, the trust gain is zero, and the aesthetic bar collapses.
Retention lifted roughly four points quarter over quarter. The main driver was the onboarding redesign that shipped mid-March, though the control cohort overlaps with a minor pricing change that may have contributed.
Loud color turns a sentence into a traffic warning. Subtle type carries the same signal without shouting.
Users already read weight as confidence.
Readers interpret typography continuously. A phrase in italic reads like a softened voice. A sentence in lighter weight reads like a whisper. The confidence gradient just makes this implicit signal align with the model's actual uncertainty. It's not a new UI; it's an alignment between the prose and its source.
Where this pattern earns its keep.
- Span-level, not sentence-level. Uncertainty often lives inside a clause, not across a whole sentence. Treat it that way.
- Type weight over color. Color hits too hard. Dimming a span with weight and opacity carries the same signal at a fraction of the visual cost.
- Legend on first use. A one-line key at the bottom of the first confidence-rendered answer. After that, trust the reader.
Confidence theater.
If your confidence gradient is driven by fake scores — a heuristic that highlights random spans to look rigorous — you've shipped confidence theater. The user will catch it the first time a confident-looking sentence turns out to be wrong. Every subsequent gradient is distrusted.
What this pattern gets wrong when it gets wrong.
- Confidence theater
- Language or typography that performs certainty beyond what the model actually has.
- Ghost citation
- A source is shown but doesn't actually back the claim, or links to a page that doesn't contain the quoted text.
Three shipping variants worth copying.
- Low-confidence phrases set at 75% opacity
- A legend tooltip the first time it appears in a session
- A 'show all uncertainty' toggle for power readers