Learn
1 learning path · 2 readings
Practical reading for designers who build with AI. Each path explains what to read, why it matters, and what to apply.
Human-AI interaction
Start with the job the user wants done. Then set rules for evidence, control, correction, and autonomy.
After this pathYou can choose an interface for an AI feature and review whether it supports appropriate reliance.
- Start with intent
Beyond Chat: 8 Core User Intents Driving AI Interaction
Intent-first framework for choosing the right workflow, interface, and success measure for an AI feature
Use this before you choose chat, a canvas, a queue, a digest, or a list.
Take with you
- Match the interface to the primary user intent.
- Define a success measure for that intent.
- Tune initiative, autonomy, transparency, tone, and risk separately.
Explore the full intent framework8 intents and 6 controls
An Index summary of eight user intents and six behavior controls. Use it to choose the right surface before you write a prompt.
Eight user intents
Each intent needs a different workflow, interface, and success measure.
- 01
Know and learn
Reduce uncertainty with structured answers, visible sources, editable scope, and clear paths into deeper evidence.
- 02
Create
Make the artifact the main surface. Support local edits, alternate drafts, version history, and undo.
- 03
Delegate
Show a plan before action, progress during action, and a clear receipt with review and recovery controls.
- 04
Oversee
Route risk and uncertainty into a review queue with evidence, context, and direct approve, reject, and edit actions.
- 05
Monitor
Turn continuous signals into low-noise digests with visible coverage, editable rules, and user-controlled timing.
- 06
Find and explore
Convert messy input into filters, ranked options, comparison tools, and a shortlist that stays available.
- 07
Play
Use presets, clear choices, session structure, and time boundaries to support safe and low-effort entertainment.
- 08
Connect
Set the relationship role and boundaries. Support continuity, safe exits, and escalation to real people when necessary.
Behavior controls
Tune these controls for the intent, the stakes, and the user.
- 01
Personalization
Set how much the system adapts to a person's data, preferences, context, and past work.
- 02
Initiative
Choose when the system waits, suggests a next step, or starts a useful action.
- 03
Autonomy
Set the distance between advice and action, with approval gates that match the possible harm.
- 04
Tone
Choose an emotional posture that fits the task, from neutral and factual to supportive.
- 05
Transparency
Decide when to show sources, assumptions, steps, confidence limits, tool use, and costs.
- 06
Risk appetite
Balance exploratory options against conservative answers that favor precision and predictable results.
- Apply the principles
39 Principles for Designing Human-AI Interaction
Applied principles for building AI interfaces with clear evidence, user control, recoverable failure, and bounded autonomy
Use this to review an AI feature before development or release.
Take with you
- Show evidence and make important output fast to check.
- Keep correction, rejection, undo, and escalation close to the result.
- Limit autonomy according to stakes, permission, and reversibility.
Explore all 39 principles39 principles
An Index summary of the source framework. Each card turns one principle into a product decision.
Probabilistic foundation
Design for inference, generation, and variable output.
- 01
Use AI for suitable work
Use AI for messy, ambiguous, generative, or synthesis work. Keep exact state changes in deterministic controls.
- 02
Treat variation as working material
Let users compare drafts, save alternatives, regenerate, and edit instead of treating the first result as final.
- 03
Match the interface to the task
Use inline help for small tasks, conversation for exploration, and plans with checkpoints for consequential work.
Expectation setting
Set a clear mental model before the first output.
- 04
State the scope and limits
Explain the available capabilities, context, accuracy, and coverage limits before the user relies on a result.
- 05
Help the user start
Use examples, templates, suggested actions, and structured inputs to remove the empty-prompt problem.
- 06
Present output as a draft
Use labels such as draft, suggestion, or review so users know that they must inspect the result.
- 07
Mark AI involvement
Distinguish generated, transformed, ranked, human-authored, and source material so attribution stays clear.
- 08
Match controls to expertise
Give novices wayfinders, experts configuration and override controls, and auditors logs with repeatable evidence.
- 09
Describe the system honestly
Give the AI a clear role without implying feelings, human judgment, or abilities it does not have.
Calibrated trust
Help reliance match the system's real reliability.
- 10
Show where claims come from
Connect important claims to their sources, tools, records, and input data.
- 11
Prefer evidence to confidence scores
Show the source passage, changed lines, or tool result instead of a number that can create false trust.
- 12
Make checks fast
Use direct links, highlights, previews, and diffs so a user can check important output at a glance.
- 13
Reject hidden objectives
Serve the stated task. Do not let engagement, retention, or sales goals quietly change the AI's behavior.
- 14
Support honest disagreement
Correct false assumptions, flag weak evidence, and show counterarguments instead of agreeing to please the user.
- 15
Credit creators and sources
Keep attribution and source relationships visible. State content limits when retrieved or licensed work shapes the output.
Transparency
Make the system understandable without making the main workflow noisy.
- 16
Answer five inspection questions
Show what happened, which inputs mattered, why this result appeared, why alternatives lost, and what can change it.
- 17
Layer the explanation
Show the shortest useful reason first. Keep assumptions, methods, sources, traces, and logs available on demand.
- 18
Show plans and traces
Show the plan before multi-step work, useful progress during it, and a receipt with review and undo after it.
Control and agency
Return control as soon as the user needs it.
- 19
Make suggestions easy to dismiss
Let users accept, ignore, edit, undo, regenerate, or revert a suggestion without breaking their current flow.
- 20
Ask only when uncertainty matters
Ask a specific question when ambiguity changes the result or raises risk. Otherwise, state the assumption and continue.
- 21
Offer local and global controls
Let users shape one result and set standing rules for memory, data access, automation, and defaults.
- 22
Respect the user's attention
Interrupt only when the value of timely help exceeds the cost of breaking focus.
- 23
Keep assistance accessible
Make edits, citations, warnings, voice flows, and traces work with keyboards, assistive technology, and different cognitive needs.
- 24
Name the rule behind behavior
Show whether a user choice, administrator policy, safety rule, privacy limit, technical constraint, or placement shaped the result.
Graceful failure
Design recovery for expected errors and uncertainty.
- 25
Limit the effect of errors
Use previews, checkpoints, history, rollback, and logs that match the possible consequences of a mistake.
- 26
Match precision to evidence
Use ranges, options, or partial answers when evidence is weak. Mark the parts that need human review.
- 27
Carry context into a handoff
Send the summary, unresolved questions, completed actions, and relevant data when a person must take over.
- 28
Make refusal a useful route
State the limit and the reason, then offer the nearest safe action that still supports the goal.
Co-creation
Keep generated work open to direct change.
- 29
Keep output editable
Support in-place edits, selected regeneration, version comparison, and continued work from the current state.
- 30
Add friction before commitment
Keep exploration fluid. Add review before a user publishes, sends, approves, or commits a consequential result.
- 31
Expose intent controls
Use examples and structured controls so users can express intent without learning hidden prompt techniques.
Responsible autonomy
Constrain action through stakes, permission, and reversibility.
- 32
Scale autonomy with risk
Automate low-risk reversible work, notify for moderate work, and require approval for irreversible or high-risk action.
- 33
Make data access revocable
Show what data the system can use and why. Ask before wider access and make removal easy.
- 34
Protect other people's privacy
Do not expose, combine, or infer sensitive details about people who did not choose the interaction.
- 35
Separate data from instructions
Keep instructions, reference data, tools, and allowed actions distinct so untrusted content cannot control the system.
Sustained reliance
Maintain useful behavior as cost, quality, models, and data change.
- 36
Explain the wait
Show meaningful stages, tool activity, and progress. Offer cancel or background options for long work.
- 37
Show cost when it changes a choice
Expose time, money, energy, rate limits, or credits when that information helps users choose an approach.
- 38
Measure healthy reliance
Measure quality, correction, and appropriate use. Do not treat acceptance, regeneration, or session length as success alone.
- 39
Prepare for model and data changes
Version important behavior, use evaluations, and test reliance again when a model or data source changes.