Beyond Accuracy: A Survey of Generalized Additive Models and the Underutilization of Their Interpretability

In high-stakes public and civic applications—such as healthcare, finance, and public services—the ability to explain a model's decisions is often a prerequisite for accountability and public trust. While the historical trade-off between interpretability and accuracy has long favored black-box models, generalized additive models (GAMs) have recently emerged as a competitive alternative, offering both high performance and inherent interpretability across various architectures. However, technical capability alone does not guarantee responsible deployment. Through a systematic literature review, we assess the evolution of GAM-based models and, critically, evaluate how researchers and practitioners are leveraging their interpretability in practice. Our findings reveal a concerning gap: while the performance of these models is steadily improving, their interpretability—the very property that makes them valuable for auditing and public trust—is frequently underutilized or superficially evaluated. We argue that for interpretable models to translate into meaningful social good, the research community must move beyond reporting performance metrics and focus on how interpretability enables external oversight, user comprehension, and the prevention of unintended harms at a population scale. We conclude with recommendations for aligning future research on interpretable models with the goals of trustworthy and accountable AI.