The feedback between planetary warming and soil carbon loss has been the focus of considerable scientific attention in recent decades, due to its potential to accelerate anthropogenic climate change. The soil carbon temperature sensitivity is traditionally estimated from short-term respiration measurements - either from laboratory incubations that are artificially manipulated or from field measurements that cannot distinguish between plant and microbial respiration. To address these limitations of previous approaches, we developed a new method to estimate soil temperature sensitivity (Q.sub.10) of soil carbon directly from warming-induced changes in soil carbon stocks measured in 36 field experiments across the world. Variations in warming magnitude and control organic carbon percentage explained much of field-warmed organic carbon percentage (R.sup.2 = 0.96), revealing Q.sub.10 across sites of 2.2 [1.6, 2.7] 95 % confidence interval (CI). When these field-derived Q.sub.10 values were extrapolated over the 21st century using a post hoc correction of 20 Coupled Model Intercomparison Project Phase 5 (CMIP5) Earth system model outputs, the multi-model mean soil carbon stock changes shifted from the previous value of 88 Â± 153 Pg carbon (weighted mean Â± 1 SD) to 19 Â± 155 Pg carbon with a Q.sub.10 -driven 95 % CI of 248 Â± 191 to -95 Â± 209 Pg carbon. On average, incorporating the field-derived Q.sub.10 values into Earth system model simulations led to reductions in the projected amount of carbon sequestered in the soil over the 21st century. However, the considerable parameter uncertainty led to extremely high variability in soil carbon stock projections within each model; intra-model uncertainty driven by the field-derived Q.sub.10 was as great as that between model variation. This study demonstrates that data integration should capture the variation of the system, as well as mean trends.