Which LIME should I trust? Concepts, Challenges, and Solutions

LIME's popularity masks serious reliability problems: While LIME is widely used to explain AI decisions in critical systems, the paper identifies fundamental challenges with fidelity (whether explanations accurately reflect how the model actually works), stability (whether the same model produces consistent explanations), and applicability across different domains—meaning users may be trusting explanations that don't reliably represent what the AI is actually doing.

No single "best" explanation tool exists yet: The paper documents numerous adaptations and enhancements to LIME proposed by researchers, but notes this growing number of variations creates confusion for practitioners trying to decide which version to use, suggesting the field hasn't yet settled on reliable solutions to LIME's core problems.

Explainability tools require careful vetting before deployment: The paper's central question—"Which LIME should I trust?"—highlights that not all explanation methods are equally trustworthy, particularly in essential systems where decisions affect people's lives. Organizations using AI explanations for accountability or compliance need to understand these tools' limitations rather than assuming explanations are automatically reliable.

The field needs better guidance for practitioners: By providing a structured taxonomy and interactive resource mapping LIME's various enhancements, the paper acknowledges that current research is difficult to navigate, suggesting responsible AI implementation requires better tools to help decision-makers select appropriate explanation methods for their specific contexts.