AI Trust & Transparency
Design interfaces where users can see into the AI's reasoning, calibrate their trust appropriately, and verify claims independently. The GLASS framework makes AI decision-making visible without overwhelming users.
Core Principle
Trust is not a boolean. Users should not "trust AI" or "distrust AI" - they should develop calibrated trust: high confidence when the AI is reliable, healthy skepticism when it's uncertain. Your job is to give them the signals to calibrate correctly.
The GLASS Framework
| Letter | Principle | Design Question |
|---|
| G | Ground in Sources | Can the user trace every AI claim back to a verifiable source? |
| L | Layer Explanations | Can the user get a 5-second answer AND a 5-minute deep dive? |
| A | Advertise Limitations | Does the interface proactively tell users what the AI is NOT good at? |
| S | Show Confidence | Can the user see how certain the AI is about each output? |
| S | Support Override | Can the user correct, override, or reject AI outputs without friction? |
The Trust Calibration Spectrum
Design for the right trust level - not maximum trust.
| Trust Level | User Behavior | Design Goal | When Appropriate |
|---|
| Over-trust (Automation Bias) | Accepts all AI outputs without checking | Introduce friction to encourage verification | High-stakes decisions (medical, financial, legal) |
| Calibrated Trust | Verifies selectively based on confidence signals | Maintain - this is the target state | Most AI interactions |
| Under-trust (AI Aversion) | Rejects AI outputs even when correct | Build trust incrementally through track record | New users, after AI failures |
Trust Erosion Events (TEEs)
A single trust violation can undo weeks of reliable performance. Design for recovery:
| TEE Type | Example | Recovery Pattern |
|---|
| Confident hallucination | AI states a false fact with no hedging | Immediately acknowledge the error class; show what changed to prevent recurrence |
| Silent failure | AI gives an answer but misses a critical constraint | Add constraint-checking signals: "I accounted for X, Y, Z in this answer" |
| Inconsistency | AI gives different answers to the same question | Surface version/context differences: "This differs from my earlier answer because..." |
| Opacity | User cannot understand why AI made a choice | Retroactive explanation: "I recommended X because of [factors]. Here's what would change if..." |
Confidence Display Patterns
The Confidence Triad
Every AI output should communicate three dimensions of confidence:
| Dimension | What It Tells the User | Display Pattern |
|---|
| Certainty | How sure is the AI about this specific output? | Color-coded badge (green/amber/red) + percentage if available |
| Basis | What evidence supports this output? | Inline citations, source cards, "Based on..." prefix |
| Scope | What does this answer cover, and what doesn't it cover? | Explicit boundary statements: "This covers X but does not account for Y" |
Confidence Display Decision Matrix
| Context | Show Numerical Confidence? | Show Color Badge? | Show Source Links? |
|---|
| Casual information lookup | No - feels clinical | Optional | Yes, inline |
| Professional decision support | Yes - precision matters | Yes | Yes, with expandable detail |
| Creative generation (writing, images) | No - subjectivity makes numbers misleading | No | Show inspiration sources if applicable |
| Code generation | Yes (test pass rate) | Yes | Link to documentation used |
| Medical/legal/financial | Yes - accountability demands it | Yes, conservative (amber default) | Mandatory, with recency indicator |
Citation Architecture
Citations are the single highest-impact trust pattern for LLM-based products.
Citation Depth Levels
| Level | What Users See | When to Use |
|---|
| L0: No citation | Raw AI output | Only for creative/casual use cases with no factual claims |
| L1: Source attribution | "Based on [source name]" | Minimum for any factual claim |
| L2: Inline citation | Numbered references linked to specific claims | Professional, research, and decision-support contexts |
| L3: Quotable evidence | Direct excerpts from sources with highlighting | High-stakes contexts where users must verify independently |
| L4: Auditable trace | Full reasoning chain + every source consulted + sources rejected | Regulated industries, compliance, legal discovery |
Citation UI Patterns
| Pattern | Implementation | Best For |
|---|
| Superscript numbers | Claim text[1] with footnotes | Long-form responses, research |
| Inline source chips | "According to WHO Guidelines 2025..." | Conversational interfaces |
| Expandable evidence cards | Collapsed by default, expand to show excerpt + link | Decision-support dashboards |
| Side-panel source viewer | Click citation, source appears in adjacent panel | Document review, analysis tools |
| Confidence-colored highlights | Text segments colored by source reliability | Professional research tools |
Explanation Layering
Different users need different explanation depths at different moments. Design explanations that telescope from simple to deep.
The 5-Second / 5-Minute / 50-Minute Rule
| Layer | Depth | Content | UI Pattern |
|---|
| 5-second | Headline | One sentence: what the AI did and its confidence | Always visible - the response itself |
| 5-minute | Summary | Key factors that influenced the output, top 3 reasons | Expandable section: "Why this answer?" |
| 50-minute | Audit trail | Full reasoning chain, all sources consulted, alternative answers considered | Link to detailed view or export |
Anti-pattern: Dumping all three layers at once. The 50-minute layer should never appear unless explicitly requested.
The "Why?" Menu
Every non-trivial AI output should support a "Why?" interaction:
| "Why?" Question | What to Show |
|---|
| "Why this answer?" | Top 3 factors that influenced the output |
| "Why not [alternative]?" | What would need to change for the alternative to be recommended |
| "What are you uncertain about?" | Specific elements with lower confidence + what additional info would help |
| "What did you ignore?" | Factors the AI considered but deprioritized, and why |
| "How would this change if...?" | Sensitivity: what inputs would flip the recommendation |
Anti-Patterns
| Pattern | Why It Fails |
|---|
| Showing confidence scores without context | "87% confidence" means nothing without a baseline. Is 87% good or bad for this task? |
| Using green for everything | If all outputs are green-badged, the badge system is meaningless. Users need contrast to calibrate |
| Burying explanations behind 3+ clicks | If users can't reach the "why" in one interaction, they won't bother |
| Making citations look like legal disclaimers | Dense, tiny-font footnotes signal "cover our liability" not "verify this yourself" |
| Explaining the model instead of the decision | Users don't care about transformer architecture. They care about "why THIS recommendation for MY situation" |
| Only explaining when wrong | If explanations only appear after errors, users associate explanation UI with unreliability |
Quick Reference
| Task | Framework Element | Key Deliverable |
|---|
| Add explainability to AI product | Full GLASS framework | Explanation layer architecture + citation depth map |
| Design confidence indicators | Confidence Triad + Display Matrix | Visual system with color, basis, and scope signals |
| Recover from trust violation | Trust Erosion Events table | Recovery flow with acknowledgment, explanation, and prevention |
| Audit an AI product for transparency | Trust Calibration Spectrum | Assessment of where users fall on the spectrum + design interventions |
| Add citations to LLM outputs | Citation Architecture (L0-L4) | Citation system matched to use case risk level |
Integration
Works with: ai-error-resilience (transparency about failures), ai-conversation-architect (confidence in dialogue), ai-safety-guardrails (transparency about content filtering), ai-feedback-loops (user corrections as trust signals), ai-ux-eval (scores the transparency dimension).
