Artificial intelligence and explainable AI in air-quality analysis: six pitfalls and practical remedies

Artificial intelligence (AI) has become ubiquitous in air-quality analytics, enabling applications ranging from day-ahead PM2.5 forecasts for urban areas to near-real-time monitoring dashboards for industrial zones (Chadalavada et al., 2025). As models grow increasingly complex, becoming black boxes, explainable-AI (XAI) methods have emerged to elucidate model predictions and support evidence-based decision-making (Tasioulis & Karatzas, 2024). However, models that perform well during training frequently fail in deployment, produce inconsistent explanations across runs, and foster misplaced confidence among stakeholders (Zhu et al., 2023). This article identifies six recurring pitfalls in AI and XAI applications for air-quality forecasting and proposes actionable mitigation strategies to improve model robustness and interpretability. Air-quality datasets exhibit strong spatial and temporal autocorrelation. Random cross-validation splits distribute correlated observations, from identical/nearby locations and adjacent time periods with similar meteorological conditions, across both training and test sets (Tang et al., 2024; Zhu et al., 2023). This data leakage allows nearly identical samples to appear in both training and test sets, producing models with artificially inflated performance metrics while masking poor generalization to new locations or future periods.