Keeping AI explanations honest when compressing models

Quantization reduces model size for deployment while keeping accuracy intact, but this work shows it silently breaks counterfactual recourse: the explanations that tell users how to change inputs to flip a model's decision. The authors formalize this problem with two metrics—Validity Drop and Counterfactual Recourse Gap—that catch failures invisible to accuracy benchmarks. Counterfactual-Faithful Quantization (CFQ) trains bit allocation to preserve recourse behavior at decision boundaries while respecting a global compression budget. Experiments on lending and criminal justice datasets show accuracy-matched baselines significantly degrade explanation stability, while CFQ maintains both prediction accuracy and recourse quality across different bit widths. This addresses practitioners deploying quantized models where explainability matters for trust and compliance.