This paper shows information-theoretic lower bounds characterizing fairness-utility trade-offs in representation learning. The work is interesting, novel and timely, and has the potential to inspire new research directions in fairness in ML.