The objective of this work is to automatically detect the use of game bots in online games based on the trajectories of account users. Online gaming has become one of the most popular Internet activities in recent years, but cheating activity, such as the use of game bots, has increased as a consequence. Generally, the gaming community disapproves of the use of bots, as users may obtain unreasonable rewards without making corresponding efforts. However, game bots are hard to detect because they are designed to simulate human game playing behavior and they follow game rules exactly. Existing methods cannot solve the problem as the differences between bot and human trajectories are generally hard to describe. In this paper, we propose a method for detecting game bots based on some dissimilarity measurements between the trajectories of either bots or human users. The measurements are combined with manifold learning and classification techniques for detection; and the approach is generalizable to any game in which avatars' movements are controlled by the players directly. Through real-life data traces, we observe that the trajectories of bots and humans are very different. Since certain human behavior patterns are difficult to mimic, the characteristic can be used as a signature for bot detection. To evaluate the proposed scheme's performance, we conduct a case study of a popular online game called Quake 2. The results show that the scheme can achieve a high detection rate or classification accuracy on a short trace of several hundred seconds.