Co-constructing Explanations for AI Systems using Provenance (2507.17761v1)

Abstract: Modern AI systems are complex workflows containing multiple components and data sources. Data provenance provides the ability to interrogate and potentially explain the outputs of these systems. However, provenance is often too detailed and not contextualized for the user trying to understand the AI system. In this work, we present our vision for an interactive agent that works together with the user to co-construct an explanation that is simultaneously useful to the user as well as grounded in data provenance. To illustrate this vision, we present: 1) an initial prototype of such an agent; and 2) a scalable evaluation framework based on user simulations and a LLM as a judge approach.