Amazon SageMaker Clarify
Bias detection and model explainability across the ML lifecycle.
❓ What is it?
A SageMaker capability that measures statistical bias in datasets (pre-training) and models (post-training), explains individual predictions using SHAP values, and evaluates foundation models for accuracy and toxicity.
💡 Why does it exist?
A model that discriminates by gender or postcode is a legal and ethical failure even at 99% accuracy. Responsible AI requires EVIDENCE of fairness and the ability to explain decisions — regulators and customers both ask.
⏱️ When should you use it?
Run Clarify before training (dataset imbalance, e.g. class imbalance across facets), after training (disparate outcomes between groups), and in production alongside Model Monitor to catch bias drift.
🗺️ Where does it fit?
It runs as processing jobs inside SageMaker, reading datasets from S3 and model endpoints you point it at; reports surface in Studio and feed model cards for governance.
🔌 How do you integrate it?
Configure a Clarify processing job with the dataset, the sensitive attribute (facet), and metrics to compute; for explainability, point it at a model to generate SHAP attributions per feature.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Clarify = bias + explainability (SHAP). If the question says "explain why the model predicted X" or "detect bias", Clarify is the answer.
- Know the two bias stages by name: pre-training (data) vs post-training (model predictions).