Are Good Explainers Secretly Human-in-the-Loop Active Learners?

Explainable AI serves a dual purpose beyond transparency: The paper reveals that when practitioners use XAI techniques to understand model predictions and collect new training data, they're essentially performing active learning—a machine learning strategy for efficiently selecting which data to label next. This means XAI tools can simultaneously improve model interpretability and data collection efficiency.

Human judgment in XAI can be mathematically formalized: The researchers provide a mathematical framework to approximate how humans contribute to the explanation-based learning process, moving beyond anecdotal observations. This formalization allows rigorous comparison between XAI-based data gathering and traditional active learning algorithms.

User studies can be replaced with simulations for efficiency: Rather than conducting expensive and time-consuming user studies to evaluate whether explanations help humans select good training data, the paper suggests this utility can be assessed through simulation. This could accelerate responsible AI research by making it more practical and cost-effective to test XAI approaches.

The workflow can be systematically designed and improved: By formalizing the end-to-end process of using explanations for active learning, the paper enables researchers to identify extensions and improvements to the workflow. This structured approach supports more intentional design of human-in-the-loop AI systems rather than ad-hoc implementations.