The deployment of machine learning methods and artificial intelligence in the context of legal decision-making will require a thorough look at the concept of variance–be it that original to human decision-makers, or that of automated systems. Human judges are, indeed, noisy in their decisions, a fact that, though deplorable from the viewpoint of individual cases, may have systemic value for the legal framework as a whole: individual variance in legal cases functions as an information-collection device that propagates, throughout the legal system, a potential lack of fit between a norm and its application, and calls for a resolution. Besides, in justice as in machine learning, randomness and variance are, increasingly, not merely a byproduct but a fundamental feature that enables these systems to avoid being trapped in sub-optimal configurations. In this context, this Note offers some reflections on whether algorithmic methods should seek to suppress or replicate variance in legal decision-making, and what principles should govern any deployment of deliberate, automated variance.