A new article shows that three commercial facial-analysis programs demonstrate gender and skin-type biases, and suggests a new, more accurate method for evaluating the performance of such machine-learning systems.