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 per-group bias metrics and prompt-stereotyping evaluation in the open-source fmeval library
Quality gaps between groups and stereotyped output belong to the fairness dimension. Prompt-stereotyping evaluation in the open-source fmeval library measures the second; the published per-group bias metrics measure the first. Bedrock evaluation jobs score accuracy, robustness and toxicity, so they do not cover either. The other dimensions cover different concerns: safety is harmful content, controllability is steering and human oversight, veracity is grounding answers in something true, and transparency is disclosing how the system works.
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 has its own controls.
| Dimension | Controls |
|---|---|
| Fairness | Prompt-stereotyping evaluation in fmeval; per-group bias metrics |
| Explainability | Citations and the retrieved passages; Model Cards |
| Privacy and security | Guardrails PII filters, blocking or anonymising; IAM, KMS, PrivateLink |
| Safety | Guardrails content filters and denied topics |
| Controllability | Human review before consequential output; feedback loops |
| Veracity and robustness | Contextual grounding checks; Knowledge Bases citations; evaluation jobs |
| Governance | Model Cards, Model Dashboard, invocation logging, CloudTrail |
| Transparency | AWS AI Service Cards; in-app disclosure |
Clarify and A2I once carried rows one and five. AWS put both into maintenance on 30 June 2026 and closed them to new customers on 30 July. Existing deployments keep running. Clarify’s evaluation engine is available on its own as the fmeval library.
Bedrock evaluation jobs score accuracy, robustness and toxicity. Bias and stereotyping measurement lands in fmeval.
Why? Naming the concern narrows what you have to build.