Transparency & Explainability

About This Module

Many AI systems reach decisions that even their builders cannot fully explain. That is a problem when the output affects a loan, a diagnosis or a prison sentence. Transparency is about being open on how a system is built, what data it uses and where it is deployed. Explainability is the ability to give a reason for a particular output that a person can follow, and interpretability is how far a human can understand the model’s internal logic. Without them it is hard to contest a decision, audit a system for bias, debug it or decide whether to trust it.

The lessons begin with a short overview of explainable AI, then move to practical tools. LIME and SHAP explain individual predictions of a black-box model by showing which features mattered most. You then hear Anthropic researchers describe mechanistic interpretability, the effort to look inside large language models, including whether a model’s stated reasoning can be trusted and why models hallucinate.

The module ends with a real debate: should we build tools that explain black boxes after the fact, or use models that are interpretable by design? Cynthia Rudin argues for the second option in high-stakes settings, because an explanation of a black box may not reflect what the model actually computes. Watch the lessons in order, then continue on to Privacy & Data Protection.