Mass conservation is essential for reliable streamflow prediction, particularly in hydrologically and climatically diverse, human-influenced watersheds. This study investigates whether the Mass-Conserving Long Short-Term Memory (MC-LSTM) model improves prediction reliability compared to the state-of-the-art LSTM. We evaluated both models on their ability to simulate streamflow using data from 51 hydrologically diverse and humaninfluenced Indian watersheds. While LSTM achieved a median Nash-Sutcliffe efficiency (NSE) of 0.71, MC-LSTM slightly improved this to 0.72. However, MC-LSTM demonstrated reduced bias and better captured high flows, despite the similar NSE values. Both models faced challenges in semi-arid and non-perennial rivers, but MC-LSTM showed marginally better performance in capturing variability under such conditions. Notably, MC-LSTM was more stable in data-scarce environments, whereas LSTM was more sensitive to training data length. Additionally, results indicated that including watershed attributes significantly enhanced performance for both models. Key factors such as the number of dams, potential evapotranspiration, clay fraction, water table depth, and catchment area correlated with the relative performance improvement of MC-LSTM over LSTM. Further analysis using a composite disturbance index (CDI) revealed that MC-LSTM shows greater performance improvement over LSTM in high-human-influenced watersheds compared to low-human-influenced watersheds, highlighting its suitability under varying anthropogenic conditions. Overall, this study provides valuable insights and highlights the advantage of physics-informed deep learning approaches like MC-LSTM in improving the reliability of streamflow predictions in human-influenced, data-limited watersheds.