Multi-task learning techniques can employ the removed redundant information to improve prediction accuracy. Which features to add to the target and/or the input during multi-task learning is still an open issue. The previous study used heuristic search methods. In this paper, a random method of genetic algorithm based multi-task learning (GA- MTL) is proposed to automatically determine the features for the input and/or the target. Experimental results on data sets from the real world show that GA-MTL is easy to use and obtains better performance than heuristic methods.