Explainable AI

Overview

This area tracks methods for making model behavior and agent decisions legible, inspectable, and communicable.

Which explanations are faithful enough to support safety claims?
When is a reasoning trace monitorable even if it is not a complete record of internal computation?
How should explanation interact with causal reasoning and strategic environments?
What explanations are useful for debugging formal constraints versus learned objectives?
When does successful agent communication fail to explain shared internal meaning?
How should post-failure explanations change perceived agency, causality, and patient harm without hiding real accountability?
How can explanations expose the safety layer’s formal reason for blocking an action without pretending to explain the whole learned policy?
When does a counterfactual explanation require a causal simulation rather than a nearby contrastive example?