Domain 4 of 5 · Chapter 2 of 2

Transparent & Explainable Models

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing AWS Certified AI Practitioner AIF-C01 premium course — no separate purchase.

Included in this chapter:

  • Transparency vs. explainability: two different promises
  • AWS tooling: explainability tools vs. transparency artifacts
  • Exam-pattern recognition: stems, tool→need mapping, traps

Interpretable models vs complex models

AspectInterpretable models (decision tree, linear/logistic regression)Complex models (deep neural nets, foundation models)
ExplainabilityInherently explainable: reasoning readable from coefficients or branchesOpaque: needs post-hoc tools like SageMaker Clarify (SHAP) to attribute predictions
Typical accuracyLower on unstructured data; strong on simple, tabular problemsHigher on images, audio, and free text where patterns are complex
Transparency artifactDocumented with a SageMaker Model Card (your own model)Model Card for your own FM/fine-tune; AWS AI Service Card for an AWS-managed service
Best forRegulated or high-stakes decisions needing defensible reason codesHigh-accuracy tasks where per-prediction opacity is acceptable

Decision tree

Must justify each individual prediction? Yes (reason codes) Interpretable model, or SageMaker Clarify (SHAP) per-feature attribution No Document YOUR model's intent, risk & metrics? Yes SageMaker Model Cards versioned record of a model you build No Assess an AWS-managed AI service you consume? Yes AWS AI Service Cards AWS-authored use cases & limits of a service No (opacity OK) Complex FM / deep model max accuracy; accept lower interpretability Transparency (what the model is) and explainability (why an output) are distinct; add a human-in-the-loop and document known limits for high-stakes decisions.
Cheat sheet

Cheat sheet

  • Transparency is disclosure about what the model is
  • Explainability answers why one prediction happened
  • A model can be transparent yet not explainable
  • Interpretability trades off against predictive performance
  • Inherently interpretable model families read their own reasoning
  • Choose interpretable models for regulated decisions
  • SageMaker Clarify gives feature attribution via SHAP
  • Clarify explains both globally and locally
  • Partial dependence plots show one feature's marginal effect
  • Clarify explains tabular, vision, and NLP models
  • SageMaker Autopilot explainability is powered by Clarify
  • SageMaker Model Cards document your own models
  • AWS AI Service Cards document AWS managed services
  • A transparency artifact never makes a black box explainable
  • Open source, open data, and licensing are transparency signals
  • Human-centered design targets the person acting on the decision
  • Honesty about limits is a core explainability principle
  • Bedrock LLM-as-a-judge returns a score and an explanation per response

Unlock with Premium — includes all practice exams and the complete study guide.

Also tested in

References

  1. Responsible AI, AWS
  2. Model explainability with SageMaker Clarify, AWS docs
  3. SHAP baselines for SageMaker Clarify, AWS docs
  4. Partial dependence plots (PDP), AWS docs
  5. Amazon SageMaker Autopilot model insights, AWS docs
  6. Amazon SageMaker Model Cards, AWS docs
  7. Responsible AI resources, AWS