Chapter 4: AI-Specific Interaction Principles

Opacity

Not being able to see or inspect AI’s reasoning behind an output
Fix: make clear what the system can do and how well

Uncertainity

Can’t reliably anticipate what it’ll do
Fix: predictability should be treated as design variable

Opacity Handling

Opacity fix #1 — Show the effect: make the algorithm's impact visible through direct comparison rather than only explaining its internals.

Model Card: standardized documentation describing intended use, limits, and performance so users/developers understand the system before relying on it.

Opacity fix #3 — Tested explanations: explanation/reasoning displays must be evaluated for whether they actually help users understand and act appropriately; simply adding more explanation is not automatically better.

Microsoft 18 Guidelines

4 categories: Initials, During Interaction, When Wrong, Over Time
G1 (make clear what the system can do)
G2 (make clear how well it can do it)
G6 (mitigate social biases)
G9 (support efficient correction)
G10 (scope services when in doubt)
G11 (make clear why the system did what it did).

8 Interaction Patterns

Prompting/re-prompting
Streaming
Confidence signaling
Citation UI
Context accumulation
Human-in-the-loop checkpoints
Correction as interaction ◦ Hedging

6 Capabilities

Capability
Calibration
Expectation
Control
Continuity
Recovery