practiceExperience & interface

Human–AI Interaction

Human–AI interaction designs how people understand, direct, evaluate, correct, and recover from systems whose behavior is adaptive or uncertain.

Download Markdown (.md)

01What it means

Human–AI interaction applies human-computer interaction practice to products whose outputs can be probabilistic, adaptive, generative, or difficult to explain. It covers expectation setting, input and feedback, uncertainty, control, correction, learning over time, disclosure, and recovery. It is broader than chat interface design and distinct from merely adding an AI capability to an existing flow.[1][2]

People and adaptive systems shape one another over time Intent, model output, human judgment, and correction form a joint decision loop

Good human–AI interaction calibrates trust, preserves meaningful control, exposes uncertainty, and supports correction and recovery across different users and stakes.

02Why it matters

Users form mental models from product behavior and may over-trust fluent output or abandon a useful system after surprising failures. Thoughtful interaction design calibrates expectations, preserves agency, supports verification, and makes errors recoverable without shifting all burden to the user.

03When to use it

Use human–AI interaction methods whenever model behavior affects a user-facing experience, even if the AI is invisible. Increase transparency, confirmation, recourse, and expert involvement with uncertainty and consequence. A conventional interaction may be better when deterministic behavior serves the task.

04How it works

  1. Research the user goal, context, capability expectations, affected people, and consequences of correct and incorrect behavior.
  2. Communicate what the system can do, what information it uses, and where its limits matter at the moment of use.
  3. Give users effective ways to steer input, inspect evidence, edit output, decline automation, and recover from errors.
  4. Design feedback and adaptation without implying that every correction retrains the model or granting silent personalization.
  5. Evaluate comprehension, calibrated trust, accessibility, workload, error recovery, and outcomes across repeated use.

05A practical example

Example

A writing assistant labels suggestions, shows changes as a reviewable diff, preserves the original, and lets users accept parts rather than the whole. When it lacks evidence for a factual rewrite, it marks uncertainty and offers source search instead of presenting certainty.

06Where it breaks

No interface can make an unreliable or unsafe model dependable. Explanations may be plausible without revealing the true cause, and repeated confirmation can create fatigue. User preferences conflict, model behavior changes, and transparency can overwhelm people if it is not contextual.

Common mistakes

  • Designing only a chat box without considering the surrounding task, evidence, action, and recovery flow.
  • Using confidence-like language or anthropomorphism that encourages inappropriate trust.
  • Hiding how users can correct, undo, opt out, or reach a person.
  • Measuring engagement while ignoring comprehension, error recovery, accessibility, and downstream harm.

07Sources

  1. Guidelines for Human-AI InteractionACM CHI Conference on Human Factors in Computing Systems · Peer-reviewed researchGuidelines for Human-AI Interaction
  2. People + AI GuidebookGoogle People + AI Research · Design guidePeople + AI Guidebook
View saved topics

Start typing to search titles, definitions, categories, and frameworks.