Influx and loss are the two most common downhole complexities. They not only cause the reservoir damage, increase the exploration cost, reduce the drilling efficiency; but also induce major malignancy. Therefore accurate and early detection of influx and loss during drilling is of great significance. Traditional influx and loss detection methods have the shortcoming of monitoring time lagging and high costs. As the rapid development of artificial intelligence techniques, researchers start to detect influx and loss using artificial intelligence method. This work adopted two machine learning algorithms(Random forests and Support vector machine) according to their characteristics to detect influx and loss during drilling in real-time. The detection methods includes four steps: 1) Generating raw influx/loss raw data set by combining real-time drilling data and drilling history data; 2) Pre-processing raw data set to obtain training data set; 3) Training classification model of random forests and SVM by training data set and algorithms; 4) Predicting influx/loss by the trained model according to the new real-time data. The case study shows that influx and loss can be detected accurately in early stage by both random forests method and SVM method after proper pre-processing the raw data and optimizing algorithm parameters. The detection accuracy of the sample data from four wells exceeds 90%. This work demonstrate a new way to detect influx and loss by utilizing huge drilling data and machine-learning algorithms, and the detection results are satisfying.