The response of soil carbon to future climate remains a major uncertainty in the global carbon cycle. Recently, machine learning approaches have become prevalent in investigations of soil organic carbon (SOC) storage and dynamics. In our studies, we applied machine learning approaches to SOC field observations (n=6,213), data of environmental factors (n=31), and three earth system model projections to (1) develop functional relationships between environmental factors and SOC stocks, (2) benchmark land surface model representations of environmental controls on SOC stocks, and (3) predict decadal changes in continental United States (U.S.) surface SOC stocks under future emission scenarios. We found that the functional relationships between environmental controls and SOC stocks derived from our analysis had similar prediction accuracy to those obtained using the random forest machine learning approach. However, current representations of environmental controls on SOC stocks in the coupled model intercomparison project phase six (CMIP6) Earth system models are not consistent with field observations. Additionally, the decadal and total SOC changes in continental U.S. surface soils predicted by our ensemble machine learning approach differed from those predicted by CMIP6 simulations. We identified specific land cover types and areas within the continental U.S. where machine learning and Earth system models both agree and disagree on the direction of SOC change under future climate scenarios. In summary, computational approaches can help in quantifying climatic impacts on SOC and reducing the uncertainty in model projections of future carbon climate feedbacks.