Guidelines for responsible AI
AWS Certified AI Practitioner exam topic, about 14% of the exam. Updated .
The Shortlist: play by chapter
Chapter 1: Fair to everyone
2 questions
Chapter 2: Show your working
2 questions
Storyline: The Shortlist
A recruitment agency uses AI to shortlist candidates and learns how to keep it fair, safe and explainable.
The agency is Kestrel Talent, with 150 people. You meet Zara (head of people and compliance), Ben (data scientist, builds the shortlisting model), Theo (senior recruiter, speaks for the candidates) and Ines (legal counsel, thinks about regulators and audits). Across two chapters, you are their advisor.
Chapter 1: Fair to everyone
Responsible AI covers bias, fairness, inclusivity, robustness, safety and veracity. A model can look accurate overall and still treat one group worse, so check outcomes group by group (subgroup analysis). Amazon SageMaker Clarify reports bias metrics for features such as age or gender, before and after training. (AWS has closed Clarify to new customers, but the exam still covers it. For foundation models, AWS now points to Amazon Bedrock Evaluations.) Removing a sensitive column isn't enough when other features act as proxies for it. Balanced, diverse, well-labelled data and human audits help too. For generative AI, Amazon Bedrock Guardrails filters harmful content, blocks denied topics and words, and masks personal data in prompts and responses. Watch the legal risks: biased outputs, intellectual property claims, hallucinations and lost trust.
Chapter 2: Show your working
A transparent model lets you see how it reaches a decision; a black-box model doesn't. Simple models are easier to interpret, and complex ones may be more accurate but harder to explain, so weigh interpretability against performance. In regulated, high-stakes decisions, explainability often wins. Amazon SageMaker Model Cards document a model's intended use, risk rating, training, evaluation results and limitations in one place for governance and audits. Design for people too: give users a way to give feedback, and be open about when AI made a decision.