Changes in land-use and land cover (LULC) occur in response to economic development and the growing demand for food. However, the impact on water availability for downstream users is often overlooked in land management policies, likely because of the lack of a well-established approach for evaluating the impact of LULC changes on river flows. This study explores the use of long short-term memory (LSTM) networks trained on data from thousands of stream gauges across the United States (US). Three LSTM networks were considered, each using different levels of LULC information: none, static (constant) and dynamic (updated annually). For this study, we created the DROMEDARY US dataset that incorporates the Cropland Data Layer dataset from the US Department of Agriculture, reflecting significant human-related LULC changes, and includes significantly more basins (3,246) than existing datasets commonly used for benchmarking hydrological models. All LSTMs demonstrated good out-of-sample prediction skills across the US. However, the ones using dynamic LULC information outperformed the others by a significant margin in reproducing observed changes in flow following changes in fallowing, an agricultural practice used to let the land rest after intense cropping cycles, or to spare water during droughts. Interestingly, using static LULC information performed worse than using no LULC information, highlighting that use of inaccurate (or outdated) information can degrade model performance at reproducing the effect of change in LULC, while having good streamflow prediction skill. Scenarios involving increased fallowed land highlighted further the benefits of using dynamic, emphasizing the need for frequently updated LULC datasets.