4.2.1. Transparent vs. Opaque Models
First Principle: AI models exist on a spectrum of transparency, from simple, easily interpretable "glass-box" models to complex, "black-box" models whose internal logic is opaque, creating a trade-off between performance and explainability.
- Transparent Models ("Glass-box" or "White-box"):
- Concept: Models whose internal decision-making process is easy for a human to understand.
- Examples: Linear Regression (you can see the weights for each feature), Decision Trees (you can follow the path of decisions in the tree).
- Advantage: High explainability, easy to debug.
- Disadvantage: Often less powerful and may not capture complex patterns as well as opaque models.
- Opaque Models ("Black-box"):
- Concept: Models that are extremely complex, making it difficult or impossible to understand their internal reasoning.
- Examples: Deep Neural Networks, Large Language Models, complex ensemble models like Gradient Boosting.
- Advantage: Often achieve state-of-the-art performance on complex tasks (like image recognition or language generation).
- Disadvantage: Low intrinsic explainability, making them harder to trust and debug. This is why techniques like SHAP feature attribution are needed to provide post-hoc explanations.
Scenario: A regulator asks a bank to explain exactly how its credit-scoring AI model works.
Reflection Question: Why would having a transparent model (like a simple decision tree) be highly advantageous in this situation compared to an opaque model (like a deep neural network), even if the opaque model is slightly more accurate?
💡 Tip: The choice between a transparent and an opaque model is a business decision. For high-stakes decisions that require regulatory approval or deep user trust, a slightly less accurate but fully transparent model might be the better choice.
