Cloud concept · Updated

Human-Centered Design - Human-Centered Design is an iterative approach to building systems around the needs, capabilities, and circumstances of the people who use them or are affected by them.

In short

For AI systems, human-centered design makes predictions and decisions understandable, useful, and open to appropriate human challenge or correction. It combines user research, accessible explanations, feedback mechanisms, and meaningful oversight so that technical performance supports real human needs rather than becoming the only measure of success.

Cloud Concepts Hub →

6 key aspects of Human-Centered Design

Human-centered AI combines research, understandable explanations, accessible interaction, actionable feedback, and meaningful oversight throughout the system lifecycle.

01
Research grounded in affected people's needs

Identify users, domain experts, and people affected by AI-assisted decisions before selecting models or workflows. Translate their goals and potential harms into requirements and measurable acceptance criteria.

02
Audience-appropriate decision transparency

Disclose the role of AI and explain relevant inputs, influential factors, limitations, and available next steps. Match the explanation to the audience and distinguish model attribution from causal evidence.

03
Actionable feedback and contestability

Provide ways to report errors, correct information, and request review of consequential outcomes. Connect each submission to an accountable handling process rather than collecting feedback without a response path.

04
Meaningful human oversight

Give reviewers sufficient context, time, and authority to reject or override automated recommendations. Define escalation rules according to uncertainty, potential harm, and operational requirements.

05
Accessible and inclusive interaction

Test interfaces and explanations with representative users, including people with accessibility or language needs. Evaluate whether different groups can understand outputs and exercise the same correction and review options.

06
Iterative evaluation and calibrated reliance

Measure comprehension, task success, error recovery, and review outcomes alongside model quality. Improve the system without encouraging users to treat confident-looking output as proof of correctness.

Cloud services for Human-Centered Design

These services support different parts of a human-centered AI application rather than providing human-centered design automatically. SageMaker AI can supply model explanations for presentation in a user interface. API Gateway can receive feedback and review requests, while Step Functions coordinates the application's handling and escalation process. Amazon Augmented AI can support configured human-review tasks. Engineers must connect these components with appropriate storage, identity controls, and user interfaces, then test explanation quality, accessibility, reviewer effectiveness, and whether feedback leads to timely resolution.

AWS Step FunctionsAWS Step Functions

AWS Step Functions can orchestrate application-defined feedback, escalation, and approval workflows. Callback patterns can pause processing while an external human-review application obtains a decision; the review interface and authorization controls must be implemented separately.

Amazon API GatewayAmazon API Gateway

Amazon API Gateway can expose application endpoints for submitting corrections, reporting problematic outputs, and requesting review. Backend code must validate submissions, associate them with the relevant decision, enforce access controls, and provide an appropriate response.

Human-Centered Design exam questions

Real practice exam questions on Human-Centered Design, with the correct answer and its explanation. Answer them with every option in the exam simulator.

AIB-C01 (beta)

A housing nonprofit is introducing AI to recommend temporary accommodation for displaced families. The goal is to move caseworkers from manually searching listings to helping families evaluate trade-offs and handling exceptional needs. In a pilot, the recommendations meet cost and availability targets, but caseworkers cannot explain why particular properties are recommended. Families also struggle to correct assumptions about accessibility and proximity to caregivers. Which action should the transformation leader prioritize before expanding the pilot?

Show answerCORRECT ANSWER

Work with caseworkers and families to prototype understandable comparisons, editable household constraints, and ways to challenge recommendations; test whether these features improve placement decisions and caseworker support.

This approach uses the experiences of both users and affected people to redesign the workflow. Understandable trade-offs and correction mechanisms let AI support the search while caseworkers apply contextual judgment and empathy. Testing decision quality establishes whether the design helps people, not merely whether recommendations meet technical targets.

AIF-C01

A retailer uses Amazon Bedrock to generate restocking recommendations from sales and inventory records. When recent inventory records are unavailable, the application uses estimated stock levels. During a pilot, store managers mistake these estimates for verified inventory counts and place unnecessary orders. Which application change BEST supports human-centered transparency?

Show answerCORRECT ANSWER

Clearly distinguish recorded counts from estimates, show data freshness and relevant limitations, and ask managers to verify estimated stock levels before ordering.

This design gives managers understandable, actionable context about the recommendation. Identifying estimates, showing data freshness, and requesting verification help users recognize limitations and correct assumptions before taking action.

Practice Human-Centered Design in a full exam

Questions on Human-Centered Design in each practice exam, with every answer option and an explanation for each:

Start a free practice exam

What is Human-Centered Design

3 min read

Human-centered design starts with people and their tasks, not with a model or cloud service. Engineers work with intended users, domain experts, and people affected by decisions to understand goals, constraints, potential harms, and the consequences of errors. In an AI application, success therefore includes whether users can interpret an output, recognize its limitations, and take an appropriate next step—not simply whether the model achieves a high accuracy score.

For explainable AI, the central design question is what a particular person needs to understand to act responsibly. A customer may need a plain-language explanation of a rejected application and instructions for correcting inaccurate information. A reviewer may need the main factors influencing that individual prediction. A model developer may need aggregate feature-importance and error-analysis results. These audiences require different explanations; exposing the same technical chart to everyone is not effective transparency.

AI decision transparency includes clearly identifying where AI is used, describing its intended purpose and limitations, and explaining how its output affects the final decision. For example, a cloud-hosted application might disclose that an eligibility recommendation is automated, show relevant input information, explain influential factors, and provide a review option. Feature attribution describes how inputs contributed to a model output under a particular explanation method; it does not establish real-world causation. A generated narrative about a decision is also not automatically a faithful explanation of the underlying model.

User-feedback mechanisms turn understanding into participation. Useful controls include reporting an incorrect answer, correcting source data, explaining why a recommendation was unhelpful, and requesting human review. In a cloud architecture, an application can send feedback through an API, store it with the relevant prediction identifier and model version, and route serious complaints to a review workflow. Collect only necessary personal information and restrict access to sensitive records. Feedback should be validated and assessed before it influences training; automatically learning from every submission can introduce manipulation, bias, or privacy problems.

Human oversight must provide meaningful authority rather than a ceremonial approval button. Reviewers need relevant evidence, adequate time, clear escalation criteria, and permission to override an automated recommendation. For example, a document-processing system can route uncertain extractions to trained staff before downstream processing. A high confidence score does not guarantee correctness, and thresholds should reflect validated performance and the consequences of mistakes. Designs should also reduce automation bias by helping reviewers evaluate evidence instead of merely accepting the system's suggestion.

Application teams apply these principles iteratively through prototypes, usability studies, accessibility testing, and production monitoring. Test whether representative users actually understand explanations, can find the appeal mechanism, and can complete their task. Examine outcomes across relevant user groups rather than relying only on average satisfaction. Support assistive technologies and language needs where appropriate.

Trade-offs include additional latency and staffing costs for human review, computation costs for explanations, and tension between detailed disclosure and protection of private data or security-sensitive logic. Layered explanations can provide a concise reason first and more detail when needed. Human-centered design does not require revealing source code, making every decision manual, or maximizing user trust. Its goal is appropriately calibrated reliance: people should understand when the system is useful, when it may fail, and how to obtain help or challenge an outcome.

Learn Human-Centered Design in lessons

2 short lessons on Human-Centered Design in 2 certification courses.

Open the lessons