What is model interpretability?

Definition

Model interpretability studies how information is represented and transformed inside an AI model. Researchers use behavioral tests, internal measurements, and explanatory methods to identify which inputs, features, or computations contributed to an output.

Interpretability can support debugging, evaluation, and safety work, but current methods provide incomplete views. An explanation may be useful without capturing the model's full internal process or reliably predicting its behavior in every new situation.

ELI5

Model interpretability is about understanding how an AI system reached an answer instead of seeing only the final result. It is like opening a machine to learn which parts affected what it did.

For example, if an AI system rejects an image, an interpretability method might show whether it focused on the object, the background, or an irrelevant mark. That clue still needs careful testing.

Frequently asked questions

Why does model interpretability matter?

It can help people diagnose errors, test safety claims, understand learned features, and decide whether a result deserves trust.

Can interpretability explain every model decision?

No. Modern models contain complex distributed computations, and current methods usually reveal only part of what influenced an output.

Videos explaining model interpretability

  1. Nick Saraev, Jack Roberts and Jakub Pachocki with AI safety and compute infrastructure themes