Due to the coarse spatial resolution of General Circulation Model (GCM) outputs, their direct application in water resources studies and climate change impact assessments is limited. Therefore, downscaling these outputs is important. This study aimed to evaluate the performance of an Artificial Neural Network (ANN) for downscaling temperature and precipitation outputs from CMIP6 climate models in the Dorudzan Watershed, Fars Province, Iran. Observed temperature and precipitation data, NCEP/NCAR reanalysis data, and outputs from eight CMIP6 climate models were used. The performance of climate models was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE), and the HadGEM3-GC31-LL model was selected as the most appropriate model for the study area. Subsequently, downscaling was performed using selected large-scale atmospheric variables and an ANN model developed in MATLAB. The results showed that the ANN model provided high accuracy in simulating monthly temperature, with an R² value of 0.99 in both training and testing stages. The RMSE values were 0.0191 and 0.0194°C for training and testing stages, respectively. For precipitation, R² values of 0.84 and 0.83 were obtained for the training and testing stages, respectively. Climate projections for 2025–2050 indicated an increasing trend in temperature and a decreasing trend in annual precipitation under all emission scenarios, with the greatest changes observed under the SSP5-8.5 scenario. Overall, the results demonstrated that ANN is an effective approach for downscaling CMIP6 climate model outputs and assessing future climate changes at the watershed scale.