Seismic behavior of gravity dams has long been evaluated using a representative two-dimensional (2D) system. Formulated for the gravity dams built in wide canyons, the assumption is nevertheless utilized extensively for almost all concrete dams due to the established procedures as well as the expected computational costs of a three-dimensional model. However, a significant number of roller-compacted concrete dams, characterized as such systems, do not conform to the basic assumptions of these methods by violating the conditions on canyon dimensions and joint-spacing/details. Based on the premise that the 2D modeling assumption is overstretched for practical purposes in a variety of settings, the purpose of this study is to critically evaluate the use of 2D modeling for the prediction of the seismic demands on these systems. Using a rigorous soil-structure interaction approach, the difference between the two and three-dimensional response for gravity dams was investigated first in the frequency domain for a range of canyon widths and foundation to dam moduli ratios. Then, the time domain differences between the crest displacements and the maximum principal stress were obtained using 70 different ground motions in order to show the possible bias introduced into the analysis results due to the modeling approach. The results of the study show that even for relatively wide canyons, the 2D analysis can lead to misleading predictions. Copyright (c) 2017 John Wiley & Sons, Ltd.