Researchers test reversible AI rules to improve existing fraud detectors

Researchers report in a preprint that rules generated by language models improved a real-world e-commerce fraud detector without changing its underlying model. The overlay gained an average +2.26% recall at equal precision and +1.60% precision at equal recall across the tested production cut-offs, while continued training caused performance to regress. The method asks language models to propose rules using features the detector underweights, screens those rules for empirical support, and applies a reversible adjustment to its scores. Pooling three models reduced the variability of using one model. On seven designed bias-shift benchmark variants, precision at equal recall improved throughout, but recall at equal precision stayed near neutral.