Interpretability is the research field that studies what is actually happening inside neural networks: identifying the internal features, circuits, and computat
The research field that studies what is actually happening inside neural networks, identifying the internal features, circuits, and computations that produce behavior. Mechanistic interpretability aims to reverse-engineer models the way one would reverse-engineer a program.
Explainability gives human-understandable reasons for a specific output; interpretability studies the model's internal mechanics themselves. Interpretability underpins safety, detecting deceptive or dangerous internal patterns, and the goal of auditing models by inspection rather than by behavior alone.