Exam-style
A review board asks how the team will detect whether their assistant gives noticeably weaker answers to some customer groups than others, and whether it repeats stereotypes about them. Which responsible-AI dimension is that, and which control carries it?
Reveal the answer
D. Fairness, carried by bias metrics and prompt-stereotyping evaluation, run as Bedrock evaluation jobs or with the fmeval library
Quality disparity between groups and stereotyped output are the fairness dimension, and bias metrics with foundation-model stereotyping evaluation, run as a Bedrock evaluation job or with the open-source fmeval library, are the controls that measure them. The other dimensions each own a different concern: safety is about harmful content, controllability is about steering and human oversight, veracity is about grounding answers in something true, and transparency is about disclosing how the system works. Mapping the concern to the right dimension is what picks the right tool.
Q. AWS names its responsible-AI dimensions. Roughly, what are they?
A. Fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance, and transparency. Each maps to concrete controls:
| Dimension | Controls that carry it |
|---|---|
| Fairness | Bias and prompt-stereotyping metrics from Bedrock evaluation jobs or the open-source fmeval library |
| Explainability | Per-answer citations and traceability; SageMaker Model Cards for the record |
| Privacy and security | Guardrails PII filters and anonymisation; IAM, KMS, PrivateLink around the data path |
| Safety | Bedrock Guardrails content filters and denied topics |
| Controllability | A human review loop (Step Functions or SQS feeding your own reviewer UI); monitoring and feedback loops to steer behaviour |
| Veracity and robustness | Contextual grounding checks, Knowledge Base citations, model evaluation jobs |
| Governance | Model Cards, Model Dashboard; invocation logging and CloudTrail for audit |
| Transparency | AWS AI Service Cards; Model Cards and visible citations in your own app |
The SageMaker names that used to carry the first and fifth rows, Clarify for bias metrics and A2I for human review, moved to maintenance in June 2026 and close to new customers from the end of July. Existing deployments keep running; Clarify’s evaluation code continues as the open-source fmeval library.
Why? The point is to map a concern to the right dimension and tool, not to hand-wave being responsible.