It has long been known that human observers can identify actions based on how people move, even from very impoverished motion depictions such as Point Light Displays (PLDs). This study investigates how humans classify actions, and what types of motion information they use to do so. Using a newly available technique (OpenPose) for extracting human joint locations from natural video, we created three types of reduced displays: PLDs, stick figures, and motion flow videos. Participants identified actions in these videos through verbal responses, and these responses were analyzed for semantic similarity using a Natural Language Processing model. A Hierarchical Bayesian Model further compared semantic similarities across video conditions. Results showed the highest intersubjective agreement (a proxy for proportion correct) for stick figures, followed by PLDs, and the lowest for motion flow videos. These results suggest that dynamic pose representations are crucial for accurate action classification, with motion flow supporting only coarse classification. The same pattern held across different action categories, such as instrumental versus locomotion and upper versus lower limb actions.