Core Principles
- Set expectations: Tell users what AI can and cannot do
- Show confidence: Indicate AI's certainty level when relevant
- Allow override: Users can always correct or override AI decisions
- Be transparent: Show how AI arrived at its output when possible
- Fail gracefully: AI mistakes should be easy to spot and fix
Patterns
| Pattern | Example |
|---|---|
| Suggestion | Email autocomplete, search suggestions |
| Classification | Spam detection, content categorization |
| Generation | Text, image, code generation |
| Prediction | Estimated delivery time, churn prediction |
| Automation | Auto-tagging, smart replies |
AI Suggestion UX
- Show suggestion inline or as a separate element
- Accept: one tap/click or keyboard shortcut
- Reject: easy dismiss, don't re-suggest the same thing
- Edit: modify the suggestion before accepting
- Feedback: way to indicate if suggestion was helpful
AI Transparency
- 'AI-generated' label on AI content
- Confidence scores when relevant (but don't overwhelm)
- Source attribution for AI summaries
- Audit trail for AI decisions in business contexts