Optimal Timing of Explanations in AI Agents

Determine when AI Agents should provide explanations—prior to executing actions, during planning, or after actions—to optimize user understanding, contestability, and the ability to reverse or correct agent operations.

Background

The paper proposes using explainable AI both post‑hoc and ex ante to improve user comprehension and actionability regarding AI Agent behaviors. While post‑hoc explanations can aid contestability, the authors argue that pre‑execution "pre‑explanations" may help users refine instructions and avoid undesired outcomes.

They explicitly note that deciding the timing of explanations is an open research area, underscoring the need for empirical and technical work to identify when explanations best support user goals and responsible agent operation.