Online car-hailing is a kind of widely used mobile application. Its core problem is to assign requests to taxi drivers with different goals. Although extensive research on task allocation has been carried out, a largely ignored problem is the income equality of drivers. Due to the short-sighted optimization and time-consuming allocation, fairness and utility receive less attention in the research on fair task allocation. In this study, an efficient task assignment scheme, learning to assign with fairness (LAF), was proposed to optimize both utility and fairness. It adopts reinforcement learning to allocate tasks holistically and proposes a set of acceleration techniques to achieve rapid and equitable allocation on a large scale. The experimental results show that the fairness, effectiveness, and efficiency of LAF are 86.7%, 29.1%, and 797% higher than the existing level, respectively.