We investigate the uncertainty associated with downscaling techniques in climate impact studies, using the Upper Beles River Basin (Upper Blue Nile) in Ethiopia as an example. The main aim of the study is to estimate the two sources of uncertainty in downscaling models: (1) epistemic uncertainty and (2) stochastic uncertainty due to inherent variability. The first aim was achieved by driving a Hydrologic Engineering Centre-Hydrological Modelling System (HEC-HMS) model with downscaled daily precipitation and temperature using three downscaling models: Statistical Downscaling Model (SDSM), the Long Ashton Research Station Weather Generator (LARS-WG) and an artificial neural network (ANN). The second objective was achieved by driving the hydrological model with individual downscaled daily precipitation and temperature ensemble members, generated by using the stochastic component of the SDSM. Results of the study showed that the downscaled precipitation and temperature time series are sensitive to the downscaling techniques. More specifically, the percentage change in mean annual flow ranges from 5% reduction to 18% increase. By analyzing the uncertainty of the SDSM model ensembles, it was found that the percentage change in mean annual flow ranges from 6% increase to 8% decrease. This study demonstrates the need for extreme caution in interpreting and using the output of a single downscaling model.