An Agent-Based Simulation of Fairness Interventions in Hiring

Abstract

Fairness interventions in hiring are often evaluated on static datasets, even though hiring decisions shape the future data on which firms train their models. We introduce a hiring-and-promotion simulation that makes this feedback explicit. In the model, firms observe biased proxy measurements, receive labels only for selected workers, and retrain on a cumulative datastore generated by their own decisions. This setup supports comparison among status-quo learning, observational debiasing, demographic-parity-style top-k selection, and repair of the measurement channel. Our results suggest that interventions with similar surface outcomes can differ substantially in how they alter the learned signal and the composition of the training data. The framework is intended as a testbed for studying fairness, dataset formation, and recoverability of bias mechanisms in dynamic organizational settings.