Explainability in Practice: Estimating Electrification Rates from Mobile Phone Data in Senegal

XAI techniques can reveal hidden biases in real-world applications: The study found that while an ML model trained to estimate electrification rates in Senegal appeared to work, explainability methods uncovered that it was biased with respect to population density—a bias that might have gone undetected without these explanation techniques.

Explainability validation on real data is still limited: Despite many technical XAI approaches existing in academic literature, this paper highlights a significant gap: there is "a lack of validation of these techniques on real-world datasets," meaning we don't yet fully understand how well these methods work outside controlled settings.

Domain expertise is essential for interpreting AI explanations: The authors emphasize that "the importance of domain knowledge to interpret explanations" cannot be overstated—technical explanations alone are insufficient; experts who understand the context (in this case, electrification patterns in Senegal) are needed to properly evaluate whether the model's reasoning makes sense.

Current XAI methods may constrain practical model design: The study identifies a critical challenge: "model design that might be restricted by currently available XAI methods," suggesting that the need for explainability could limit which AI approaches practitioners can use to solve real problems.

Data processing challenges complicate real-world XAI implementation: Beyond model explanation itself, the authors found that "data processing" was a major practical hurdle, indicating that responsible AI requires attention to how data is prepared, not just how models are explained.