Robustness is of central importance in machine learning and has given rise to\nthe fields of domain generalization and invariant learning, which are concerned\nwith improving performance on a test distribution distinct from but related to\nthe training distribution. In light of recent work suggesting an intimate\nconnection between fairness and robustness, we investigate whether algorithms\nfrom robust ML can be used to improve the fairness of classifiers that are\ntrained on biased data and tested on unbiased data. We apply Invariant Risk\nMinimization (IRM), a domain generalization algorithm that employs a causal\ndiscovery inspired method to find robust predictors, to the task of fairly\npredicting the toxicity of internet comments. We show that IRM achieves better\nout-of-distribution accuracy and fairness than Empirical Risk Minimization\n(ERM) methods, and analyze both the difficulties that arise when applying IRM\nin practice and the conditions under which IRM will likely be effective in this\nscenario. We hope that this work will inspire further studies of how robust\nmachine learning methods relate to algorithmic fairness.\n