The most recent climate model comparisons provide strong consensus that the climate is changing, but impacts on the terrestrial hydrologic system remain uncertain. Of particular concern is the fate of subsurface water resources in moisture-limited regions. Although average precipitation is expected to increase with a warmer climate, local precipitation amounts are likely to decrease in mid-latitudes, in regions that are dry. In recent years, a number of studies have tried to predict future groundwater recharge rates using climate model predictions. Some have acknowledged the uncertainty in future meteorological forcing by considering various climate model projections. However, uncertainty in other recharge control factors is seldom considered, even though recharge simulations are very sensitive to soil and vegetation conditions, especially in dry environments. This study presents a probabilistic approach to examining future impacts on recharge that fully accounts for uncertainty in the weather and the land-surface. It is demonstrated for a semi-arid agricultural site in the Southern High Plains, Texas with a range of global climate model predictions including both rainier and drier outcomes. An ensemble-based importance sampling approach was used to estimate conditional probability distributions for uncertain soil and vegetation parameters. These distributions were conditioned on chloride concentration and soil moisture data. The soil parameters distributions were used to derive conditional recharge probability distributions for different climate scenarios. Results for the test site indicate that deep percolation past the root zone occurs episodically, with a small number of intense events comprising a disproportionately high fraction of the long term recharge. They also suggest that future increases in total rainfall or intensity during key months would increase the probability of high magnitude percolation events. Generally, uncertainty in long term recharge predictions scale with precipitation amounts, with the higher moisture input amplifying the uncertainty from land-surface properties. This work shows that probabilistic forecasts provide a more informative characterization of uncertain soil moisture dynamics than deterministic predictions, especially when non-linearities and intermittent forcing are important.