Activation mapping in animal models of atrial fibrillation (AF) has shown that activation patterns can repeat for several cycles and then be followed by different changing or repetitive patterns. Subsequent clinical studies have suggested a similar type of activity in human AF. Our purpose was to investigate whether a computer model of human AF can reproduce this behavior. We used a three-dimensional model of the human atria consisting of 0.2-mm volumetric elements with a detailed representation of bundle structures and fiber orientations. Propagating activation was simulated over 9 seconds with a monodomain reaction-diffusion model using human atrial membrane dynamics. AF was induced by rapid pacing. Pattern recurrence was quantified using a similarity measure based on transmembrane voltage at 1000 uniformly distributed points in the model. Recurrence plots demonstrated a continuous evolution of patterns, but the speed of evolution varied considerably. Groups of upto 15 similar cycles could be recognized, separated by rapid changes. In a few cases truly periodic patterns developed, which were related to macroscopic anatomical reentry. The drivers of non-periodic patterns could not be clearly identified. For example, a spiral wave could appear or disappear without an obvious impact on the similarity between cycles. We conclude that our model can indeed reproduce strong variations in similarity between subsequent cycles, but true periodicity only in case of anatomical reentry.