GLocalX -- From Local to Global Explanations of Black Box AI Models

Black box AI models sacrifice transparency for accuracy: The paper highlights that while complex AI systems like deep neural networks and ensemble models achieve remarkable accuracy on difficult tasks, they often function as "black boxes" that humans cannot understand—a critical concern for applications in sensitive domains like healthcare, criminal justice, and finance.

Local explanations can be aggregated into understandable global models: GLocalX addresses the transparency problem by starting with local explanations (decisions for individual cases) and hierarchically combining them into simpler, interpretable global models that can emulate or potentially replace the original black box system.

Simpler models can match complex AI performance: The experiments demonstrate that GLocalX can accurately replicate the behavior of sophisticated black box models using "simple and small models," suggesting that transparency and accuracy need not be mutually exclusive—a key finding for responsible AI deployment.

The method works even with limited data access: GLocalX was tested in constrained settings with limited or no access to training data or local explanations, making it practical for real-world scenarios where organizations may have privacy concerns or data restrictions that prevent full model inspection.

Interpretability enables better accountability: By creating an interpretable layer on top of black box models, this approach supports responsible AI practices by making it possible to understand, audit, and potentially challenge AI decisions in critical applications—essential for building public trust and ensuring fairness.