An explanation only helps if the person reading it can understand it and act on it. Human-centered design for explainable AI means designing an AI system around the people who use it and the people it affects: telling them they're dealing with AI, showing why it reached an answer and how sure it is, letting them push back, and keeping a person in charge of decisions that matter. This lesson covers those principles, which the exam tests through user-experience scenarios.
Design Around the User Journey
AWS's Responsible AI guidance suggests mapping the whole journey a user takes with an AI system, because each stage needs something different:
- Before the interaction: help users learn what the system can and can't do. Where appropriate, AWS's service cards encourage disclosing to end users that machine learning is being used.
- During processing: be open about what the AI is doing, and show progress for anything slow.
- After the interaction: let users understand the output, challenge it and give feedback.
Different user groups may be affected differently at each stage, so the design should include safeguards for the groups most at risk.
Make Decisions Understandable
People affected by an AI decision should be able to see the main reasons behind it, in language suited to them. A data scientist may want detailed feature scores, while a customer needs a plain sentence such as "your application was declined mainly because of a short credit history". For generated answers, citations to source documents let users check claims for themselves. Explanations should be tested with real users, because an explanation that confuses people, or sounds more certain than it should, can do more harm than none.
Help Users Trust the Right Amount
The goal isn't maximum trust; it's appropriate trust. Users who over-trust an AI accept its mistakes, and users who under-trust it ignore good advice. Confidence indicators help: showing how certain the system is tells users when to double-check. AWS also recommends testing whether users can actually tell when to rely on an output and when to question it, and checking that confidence scores really track accuracy.
Feedback and Human Control
A user-feedback mechanism, such as a rating button, a "report a problem" link or a way to correct an answer, gives the team real-world evidence of where the system fails, and gives users a voice. Feedback should flow back into improving the system.
For high-stakes or sensitive decisions, such as those affecting someone's rights or access to essential services, AWS's guidance is to keep human review in the workflow. A clear path to a person, whether to appeal a decision or to take over when the AI is unsure, keeps people in control of outcomes the AI shouldn't settle alone.
On the Exam
| Scenario need | Design principle |
|---|---|
| Users don't realize they're talking to a bot | Disclose AI use and its limits up front |
| Users can't tell why a decision was made | Plain-language reasons or source citations |
| Users accept wrong answers without checking | Confidence indicators, and human review when confidence is low |
| The team doesn't know where the AI fails in practice | User-feedback mechanism |
| A decision affects someone's rights | Human review and a way to appeal |