Principle 01
Transparency without overload.
Reveal what a reader needs, when they need it. Every claim should be inspectable — none should be shouted. Density is not the same as depth.
AASHIKA PAREKHInteraction · Finance · AI
Designing Trust in AI
Financial Research
Invisible Confidence
How might we design AI interfaces that quietly earn trust instead of demanding it?
§ 01The problem
Analysts use AI tools to make high-stakes financial decisions.
But every time AI responds instantly with an answer, the user's real thought is:
“Can I trust this?”
That moment of hesitation is the core design problem.
Aphorism 001Speed without trust is noise.
Simulated exchangelatency 0.3s
AIHuman · internal
Revenuegrew18.4%YoY,outpacingconsensus.
↓
Where did that number come from?
hesitation
§ 02The framework
A framework
Invisible Confidence
Definition
Invisible Confidence is the collection of subtle interface signals that reduce uncertainty without interrupting workflow.
§ 03Six micro-interactions
Trust compounds one gesture at a time.
Each of the following is a fully-rendered pattern, not a screenshot. Hover, drag, and expand. Notice how the interface admits what it does not know.
AI answer · hover any figureConfidence 0.82
NVIDIA's data-center revenue grew in the last quarter, driven primarily by with a projected gross margin of .
3 sources · 2 primary · 1 model estimate
Confidence dial · drag72%
Directionalsignal
Two independent sources agree. Model interpolation within recent history.
§ 04Interactive experiments
Try it. See how trust changes.
Four small prototypes. Each one lets you interrogate an AI answer the way an analyst would — pulling threads, toggling perspectives, watching the reasoning move.
Exp 01
interactiveConfidence Evolution
Click to reveal deeper layers
layer 01 · Claim
AI answer: “NVIDIA data-center revenue grew ~154% YoY in Q2 FY25.”
Exp 02
interactiveProgressive Disclosure
Hover for sources · click for reasoning
Summary
Apple services revenue is at an all-time high, growing double-digits YoY.
hover the summary
Exp 03
interactiveConflicting Signals
Toggle to see both perspectives
Data-center capex is a decade-long build cycle.
- Hyperscaler capex guidance revised up three quarters in a row.
- Backlog visibility extends into late FY26.
- Software attach improving unit economics per GPU.
Both are true. Trust grows when the interface shows the tension.
Exp 04
interactiveVersioned Intelligence
Move between revisions to see what changed
May 18
Margin stabilizes near 18%; upside depends on Energy segment mix.
What changed
Now sourced to primary 10-Q. Removed extrapolation. Confidence up.
§ 05The full interaction
A research question, answered without pretending to be certain.
PROTOTYPE
01 · Ask
Ask a research question…
Session posture
The assistant is instructed to prefer primary filings, cite every numeric claim, and refuse when it lacks evidence.
02 · Answer
§ 06Design principles for trust in AI
Four rules we return to.
Principle 02
Intelligence should be inspectable.
A black box is a liability. An interface earns authority by exposing the shape of its own reasoning — the sources, the discarded paths, the caveats it kept.
Principle 03
Users should stay in flow while verifying.
Verification is not a detour. Hover, expand, reversibility — the reader glances sideways at the evidence and keeps reading the sentence they were in.
Principle 04
Trust is built through repetition, not claims.
No badge, no seal, no adjective. Trust is the residue of small consistencies — an honest refusal, a citation that resolves, an undo that always works.
Invisible confidence is not about making AI feel certain.
It is about making users feel equipped to decide.
Try earning your confidence.→
Research artifactNOT A PRODUCT