COURSE 26L1100% FREE
Verified 2026-08-14

Human-AI interaction basics

**Human-AI interaction (HAI)** designs how people and imperfect, probabilistic AI systems work together: when to show uncertainty, when to ask for confirmation, and how to recover from errors. Microsoft’s Guidelines for Human-AI Interact...

What it is

Human-AI interaction (HAI) designs how people and imperfect, probabilistic AI systems work together: when to show uncertainty, when to ask for confirmation, and how to recover from errors. Microsoft’s Guidelines for Human-AI Interaction (CHI 2019; HAX Toolkit) synthesize 18 evidence-based practices across initial use, during interaction, when wrong, and over time. Google PAIR’s People + AI Guidebook translates similar ideas into product patterns and case studies.

<!-- IMAGE: person ↔ AI with labels: expect, explain, control, recover -->

MEDIUM PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

HAI loop: Human intent → UI → Model → Presentation → Human interpretation → Action in world. Failure smudge 'automation bias'.

Educational Focus: Frames product AI as interaction, not just model quality.

Why it matters

A strong model with a confusing UI still fails users. Stanford HAI conversations emphasize designing at user, community, and society levels—and measuring what *people can do with* models, not only model accuracy (Amershi and others at HAI conferences).

How it works (plain)

  • Make the AI’s role clear (draft vs decide)
  • Show sources when claims matter
  • Require confirms for irreversible actions
  • Provide easy undo and handoff to humans
  • Avoid dark patterns that hide automation
  • Plan for inevitable errors (HAX Playbook mindset for NLP failures)

Everyday example

Spellcheck suggests; it shouldn’t silently send the email.

Try it

Sketch a one-screen UI for an AI feature with: role label, confidence/source, confirm, undo. Map each element to a HAX guideline number if you can.

Myths

⚠️ Myth: More autonomy is always better UX.
✓ Reality: Autonomy without visibility destroys trust.
⚠️ Myth: Anthropomorphic chatter equals good design.
✓ Reality: Clarity beats cosplay.
⚠️ Myth: Academic HCI is irrelevant to shipping.
✓ Reality: Industry toolkits (HAX, PAIR) are HCI research made operable.

Sources