When Algorithmic Nudging Backfires: Evidence From a Field Experiment Leveraging Causal Machine Learning

We study the effectiveness of an algorithmic nudge in the form of algorithmic financial advice on retirement-saving decisions. Using a field experiment (N = 4,322 customers), we compare an algorithmic nudge, that is, advice generated by a machine-learning model, with a one-size-fits-all nudge based on average historical investments. We then use causal machine-learning techniques to study heterogeneous treatment effects of algorithmic nudging. Our findings reveal substantial heterogeneity. While algorithmic nudges improve outcomes for individuals with managerial backgrounds, they backfire for those with computer science backgrounds. We further show that a targeted strategy that delivers algorithmic nudges only to individuals most receptive to them could increase retirement savings contributions by up to 34%. This research contributes to the literature on algorithmic nudging and algorithmic management.