| name | AI Error Resilience |
| description | Design graceful failure experiences for AI products - hallucinations, uncertainty, wrong outputs, and edge cases. Use when: AI hallucination UX, error handling for AI, uncertainty design, graceful degradation, AI failure recovery, confidence thresholds, safe fallbacks. |
AI Error Resilience
Design AI products that fail gracefully, communicate uncertainty honestly, and help users recover without losing trust. The RECOVER framework treats AI errors as a design material, not a bug to hide.
Core Principle
Traditional software has bugs. AI has probabilistic outputs on a spectrum of correctness. You cannot design AI UX using binary error/success patterns. Instead, design for a continuum: right, mostly right, partially right, uncertain, wrong, and dangerously wrong.
The RECOVER Framework
| Letter | Phase | Design Question |
|---|
| R | Recognize | Can the system detect when its output may be unreliable? |
| E | Express Uncertainty | Does the interface clearly communicate degrees of confidence to the user? |
| C | Contain Blast Radius | If the AI is wrong, what's the worst that can happen? How is damage limited? |
| O | Offer Alternatives | Does the user get a Plan B when Plan A might be wrong? |
| V | Verify Collaboratively | Can the user easily check, correct, or confirm the AI's output? |
| E | Evolve from Errors | Does the system learn from this error type to prevent future occurrences? |
| R | Restore Confidence | After a failure, how does the product rebuild the user's willingness to try again? |
AI Error Taxonomy
Not all AI errors are created equal. Each type requires a different UX response.
| Error Type | Description | Severity | UX Response Pattern |
|---|
| Confident hallucination | AI invents facts and states them as truth | Critical | Citation requirement + verification prompt |
| Stale knowledge | AI references outdated information | High | Timestamp + "knowledge cutoff" indicator |
| Context misread | AI misinterprets the user's intent or context | Medium | Reflect understanding before answering |
| Partial answer | AI addresses some but not all parts of a query | Medium | Checklist showing what was/wasn't addressed |
| Formatting error | Content is correct but presented poorly | Low | Easy reformat/regenerate affordance |
| Refusal overreach | AI refuses a legitimate request due to overly strict safety filters | Medium | Explain why + offer alternative path |
| Confidence inversion | AI is most confident when it's most wrong | Critical | Mandatory human review for high-stakes outputs |
The Error Severity Matrix
Use this matrix to determine the appropriate UX response based on error probability and consequence:
| Low Consequence | Medium Consequence | High Consequence |
|---|
| High Probability of Error | Auto-correct silently + log | Warn before action + suggest alternatives | Block action + require human approval |
| Medium Probability | Show confidence indicator | Present with verification prompt | Require explicit confirmation + evidence |
| Low Probability | No intervention needed | Subtle confidence signal | Add verification step for critical outputs |
Uncertainty Communication Patterns
The Hedging Spectrum
From most to least uncertain, calibrate AI language:
| Confidence Level | AI Language Pattern | Visual Signal | Example |
|---|
| 90%+ | Direct statement | Green / no indicator | "The meeting is scheduled for 3pm." |
| 70-90% | Qualified statement | Amber indicator | "Based on the data available, the meeting appears to be at 3pm." |
| 50-70% | Explicit uncertainty | Amber + explanation | "I found conflicting information. It may be 3pm, but some sources say 2pm." |
| 30-50% | Presented as possibilities | Red indicator | "I'm not confident about this. Here are the possibilities I found: [list]" |
| Below 30% | Deferred to user/human | Red + escalation | "I don't have enough information to answer reliably. Here's what might help: [resources]" |
Anti-pattern: Using the same hedging language for everything ("I think..." prepended to every response). This makes hedging meaningless. Reserve it for genuine uncertainty.
The Hallucination Defense Playbook
Prevention Layer (Before Output)
| Strategy | How It Works | Implementation |
|---|
| Source grounding | Only make claims traceable to provided sources | RAG architecture with citation requirements |
| Constraint declarations | AI states what information it's working from | "Based on the 3 documents you uploaded..." |
| Confidence gating | Suppress outputs below confidence threshold | Set per-use-case thresholds (e.g., medical = 90%, casual = 50%) |
| Scope framing | AI explicitly states what's in and out of scope | "I can help with X. For Y, you'll need [alternative]." |
Detection Layer (During Output)
| Strategy | How It Works | Implementation |
|---|
| Self-consistency checking | Generate multiple responses; flag divergence | If 3 generations disagree, show the disagreement to the user |
| Factual anchoring | Cross-reference claims against known data | Highlight claims that don't match provided sources |
| Temporal markers | Flag information that may be outdated | "This information is from [date]. Verify current status." |
Recovery Layer (After Error)
| Strategy | How It Works | Implementation |
|---|
| One-click correction | User marks output as wrong with minimal effort | Thumbs down + "What was wrong?" dropdown |
| Regeneration with guidance | User can ask AI to try again with a hint | "Try again, but focus on [specific angle]" |
| Human escalation | Seamless handoff to human when AI fails | "Let me connect you with someone who can help with this." |
| Error acknowledgment | AI explicitly owns the mistake | "You're right - my previous answer was incorrect. Here's the corrected version." |
Safe Fallback Design
Every AI output needs a fallback plan. Design the fallback hierarchy:
| Priority | Fallback Level | What Happens | User Experience |
|---|
| 1 | Graceful degradation | AI provides a less detailed but more reliable answer | "I can give you a general answer, but for specifics you'll want to check [source]." |
| 2 | Transparent limitation | AI explains what it can't do and why | "I don't have access to real-time data for this. Here's what I know as of [date]." |
| 3 | Alternative path | AI suggests another way to accomplish the goal | "I can't do X directly, but here's how you might approach it using Y." |
| 4 | Human handoff | AI transfers to a human with full context | "Transferring you to a specialist. I've shared our conversation so you won't need to repeat yourself." |
| 5 | Graceful exit | AI acknowledges the dead end respectfully | "I've reached the limit of what I can help with here. Here are some resources that might help: [links]" |
Anti-pattern: The cliff-edge fallback - AI works perfectly until it doesn't, then shows a generic "Something went wrong" error with no path forward.
The Blast Radius Audit
For every AI feature, map the blast radius of failure:
| Question | What to Document |
|---|
| What's the worst output the AI could produce? | Specific harmful scenarios, not abstract risks |
| Who is harmed if the AI is wrong? | End user, downstream consumers of the output, third parties |
| Is the harm reversible? | Can the user undo, correct, or recover from a bad output? |
| How quickly is the error detectable? | Immediately (user sees it), delayed (downstream impact), or hidden (user doesn't know)? |
| What safeguards exist? | Human review, verification prompts, confidence gates, output constraints |
Anti-Patterns
| Pattern | Why It Fails |
|---|
| Hiding uncertainty behind confident language | Creates automation bias; users can't calibrate trust |
| Showing a spinner then delivering a hallucination | Users equate "thinking time" with reliability - longer processing implies more trustworthy answers |
| Generic error messages ("Something went wrong") | Tells the user nothing actionable. Always specify: what failed, why, and what to do next |
| Requiring users to detect AI errors themselves | Users are not QA testers. Build detection into the system |
| Punishing users for reporting errors | If the correction flow is cumbersome, users will stop reporting and start silently distrusting |
| Overcorrecting after one error | Dropping confidence displays to 0 after one mistake makes the system unusable. Recalibrate, don't overreact |
Quick Reference
| Task | Framework Element | Key Deliverable |
|---|
| Design error handling for AI product | Full RECOVER framework | Error type map + severity matrix + fallback hierarchy |
| Audit AI product for failure risks | Blast Radius Audit | Risk map with harm scenarios and safeguards |
| Design hallucination prevention | Hallucination Defense Playbook | 3-layer defense (prevention, detection, recovery) |
| Calibrate uncertainty language | Hedging Spectrum | Language guidelines matched to confidence levels |
| Design safe fallback flow | Safe Fallback Hierarchy | 5-level fallback chain with user experience specs |
Integration
Works with: ai-trust-transparency (how errors affect trust), ai-conversation-architect (error recovery in dialogue), ai-safety-guardrails (preventing harmful errors), ai-feedback-loops (learning from error reports).
