Security and governance

Interpretability

stable definition
Machine-readable Download Markdown

Definition

Interpretability is the degree to which a person can understand how a model represents information or produces behavior. The target might be a single prediction, the role of a feature, a learned concept, a circuit inside a neural network, or the model's behavior across an entire class of inputs.

An interpretation should make a claim that can be tested. Feature attribution, probes, activation analysis, causal interventions, and simpler surrogate models offer different evidence and carry different limits. A visualization that looks meaningful is not enough if changing the alleged mechanism does not change the behavior.

Distinguish it from nearby terms

Explainability usually concerns a human-usable account for an output or action. Interpretability can go deeper into the model's underlying representations and mechanisms. Post-hoc explanations may approximate behavior without showing what the original model actually used.

Check your understanding

A probe can predict whether an internal activation encodes sentiment. That shows the information is available there, not that the model uses it to choose the answer. A stronger interpretation would test the causal link.