Advance warning of streamflow drought would enable improved water resource management. We used two artificial intelligence models and two reference models to forecast when streams experience drought at more than 3,000 streamgages within the contiguous United States. We forecast up to 13 weeks ahead, though we generally found 4 weeks to be the limit of performance for predicting the occurrence of severe droughts. Our best-performing model was a long-term short-term memory neural network architecture. It was able to correctly forecast the onset of drought 18% of the time, which, while low, exceeded the reference models. Our forecasts were also evaluated for the properties of the uncertainty estimates; we found the best-performing model to also have the most informative uncertainty estimates. This work highlights the challenges and opportunities to further advance hydrological drought forecasting and supports an experimental operational streamflow drought assessment and forecast tool.