Transparency discloses that AI is used and how data flows. Explainability helps humans understand why a specific decision occurred—critical for appeals, debugging, and trust.
Layers of transparency
- UI disclosure: "AI-generated" or "AI-assisted ranking"
- Model cards: intended use, limits, evaluation slices
- Data documentation: sources, refresh cadence, known gaps
Explainability methods (preview)
- Intrinsic — linear models, shallow trees show feature weights
- Post-hoc — SHAP, LIME approximate complex models
- Example-based — similar past cases influenced outcome
Explanations must be validated—plausible stories can mislead.
User-facing copy
disclosure = (
"Suggestions are ranked by an AI model using your "
"reading history. You can hide genres or reset preferences."
)
print(disclosure[:60], "...")Practice: Reflect on ethics scenarios in writing—no code required. Optional snippets illustrate policy checks only.
Important interview questions and answers
- Q: Explainability vs interpretability?
A: Interpretable models are understandable by design; explainability tools approximate black boxes. - Q: When explanations required?
A: Regulated credit, hiring, healthcare—varies by jurisdiction.
Self-check
- List three transparency layers.
- Why validate post-hoc explanations?
Tip: UI disclosure beats buried model cards users never see.
Interview prep
- Model card?
- Structured summary of intended use, data, metrics, and limitations.
- Post-hoc explainability risk?
- Plausible explanations may not reflect true model reasoning.