As AI models become more powerful, governance questions become critical: Does the model treat all groups fairly? Can its decisions be explained? Can humans intervene when needed? AIP-C01 tests how you implement responsible AI principles using AWS services.
AWS Responsible AI Principles
AWS defines six responsible AI principles: Fairness, Explainability, Privacy, Security, Transparency, and Safety. Each maps to specific AWS services — Clarify for bias detection and explainability, Guardrails for safety, Model Cards for transparency, A2I for human oversight.
!AWS Responsible AI 6 core principles
Types of Bias
Key bias types tested on AIP-C01:
: Training data doesn't represent the deployment population : Data collected differently across groups : Labels from historically discriminatory decisions reinforce existing patterns : Combining diverse groups loses subgroup-specific patterns : Model deployed in a context different from where it was trained
SageMaker Clarify — Bias Detection and Explainability
Clarify measures pre-training data bias and post-training model bias separately. Key metrics include Class Imbalance (CI), Difference in Positive Proportions in Labels (DPL), and Disparate Impact (DI).
For explainability, Clarify uses SHAP values to quantify each feature's contribution to individual predictions (local explanation) and overall model behavior (global explanation). This is essential for regulatory contexts where decisions must be explained — such as loan rejections or medical risk assessments.
Fairness Metrics
Common fairness metrics each capture a different notion of fairness:
: Equal positive prediction rates across groups : Equal True Positive Rates for actually qualified individuals across groups : Predicted probabilities match actual outcome frequencies
These metrics are mathematically incompatible in many scenarios — choosing which to optimize is a business and ethical decision, not just a technical one.
SageMaker Model Cards
Model Cards document a model's purpose, training data, performance metrics (overall and per-subgroup), bias assessment results, intended uses, and prohibited uses in a standardized format. Integrate Model Cards as mandatory artifacts in your model review and approval governance process.
Amazon Augmented AI (A2I)
A2I provides managed Human-in-the-Loop workflows. Configure Human Review Workflows with trigger conditions (e.g., confidence below 70%), assign Work Teams (internal staff or Mechanical Turk), and define Task UIs. Integrates natively with Textract and Rekognition; custom models connect via API. Essential for regulatory requirements mandating human review of high-stakes AI decisions.
Compliance Governance
AWS Config tracks SageMaker resource configuration changes and evaluates policy compliance continuously. AWS Audit Manager automatically collects evidence for compliance frameworks (GDPR, HIPAA, ISO 27001). SageMaker ML Lineage Tracking records the full chain from training data source through preprocessing to model artifact to endpoint — enabling traceability and incident response when data quality issues arise.