In this paper we propose the use of non-persistent CSMA as an anti-collision procedure for RFID active tags. Current proposals for both passive and active tags are based on the framed slotted ALOHA protocol, which does not scale well requiring additional procedures for frame length adaptation. However, active RFID devices already include carrier sense capabilities with no additional cost and, thus, CSMA may be employed seamlessly. Nevertheless, selecting the contention micro-slots of CSMA in the classical way (i.e., with a uniform distribution and an exponential back-off algorithm) does not result in an efficient identification process, as we will demonstrate. Fortunately, better choices can be found. Recently, an optimal distribution for the selection of micro-slots for event-driven sensor networks has been computed, as well as a practical implementation: the Sift distribution. In this work we propose the application of the quasi-optimal Sift distribution along with CSMA for active tag identification. By means of an analytical study, we evaluate the average time needed for identification with this mechanism and compare it with the current ISO 18000-7 and EPC “Gen 2” standard. The results reveal that the Sift-based non-persistent CSMA outperforms both of them. Moreover, it also scales much better, without the need for further adaptation mechanisms.