Guidelines For The Choice Of The Baseline in XAI Attribution Methods

Baseline choice significantly impacts AI explanations: The paper demonstrates that selecting a "baseline" (a neutral reference point used to explain AI decisions) is not a technical detail but a fundamental choice that directly determines what explanations an AI system produces, making it critical for trustworthiness.

Current practice lacks clear guidelines: The research reveals that existing explainability methods often leave baseline selection ambiguous or arbitrary, which undermines efforts to make AI systems more transparent and trustworthy to users and stakeholders.

A practical solution exists: The authors propose using "decision boundary sampling"—a method that systematically searches for the best baseline by focusing on the boundary between different AI predictions—offering organizations a concrete, implementable approach to improve explanation quality.

Better baselines enhance reliability claims: By providing clear guidelines for baseline selection, the paper helps reduce ambiguity in AI explanations, which the authors argue directly strengthens claims about whether AI models are reliable, fair, and trustworthy—key concerns for responsible AI deployment.

Explainability requires methodological rigor: The work emphasizes that responsible AI isn't just about having explanation tools, but about making deliberate, justified choices in how those tools work, suggesting that trustworthy AI requires attention to technical details that might otherwise be overlooked.