Deep learning (DL) is designed to return objective results, yet unintended inequities may perpetuate through the model. One cause is the presence of systematic social bias (often unconscious) introducing a layer of subjectivity in what was assumed to be impartial data. We use reinforcement learning (RL), a form of DL, to detect and correct for social bias. Given state information, RL makes decisions and receives quantitative feedback that it evaluates without user input. RL learns from experience and develops a set of strategies to govern its decisions. These strategies evolve as RL gathers more data. In theory, an RL algorithm fed biased data would recognize and adjust for a mismatch between expected and observed results. We simulated data where one group had systematic bias, which increased misclassification. Our RL algorithm successfully recognized and corrected for the bias, both with group id and, importantly, with an indirect measure of group id. As prediction machines, such as DL, become increasingly popular, it is vital to correct for rather than propagate social bias. RL may provide a tool to identify and correct bias in both RL and potentially DL systems.