The net carbon dioxide (CO₂) flux from land use and land-use change (F_LUC) is a major driver of anthropogenic climate change and central to mitigation strategies for achieving global emission reduction targets. Despite its importance, estimates of F_LUC are characterized by large uncertainties. In models quantifying F_LUC, the spatial distribution of potential natural vegetation (PNV) is a key component, but its influence on F_LUC estimates has not been systematically quantified. Here, we address this gap by combining pollen-based biome reconstructions and observation-based datasets of environmental conditions with machine learning to derive global PNV maps. Compared to existing PNV maps, our approach improves the representation of biome-specific spatial heterogeneity and provides sensitivity maps for quantifying how assumptions about potential forest and grassland distribution propagate into F_LUC estimates. Implementing the PNV maps as Plant Functional Types (PFTs) in the bookkeeping model BLUE, we find global cumulative F_LUC for the period 1850–2023 to be 16 % (6–27 %) higher than the default estimate. Differences at the regional scale are often even larger. Our results demonstrate that uncertainties in PNV distribution represent a substantial and previously overlooked source of uncertainty in F_LUC estimates, comparable in magnitude to other key sources. Accurate PNV mapping is therefore essential for robust F_LUC estimates, particularly at regional scales, which are required for understanding the global carbon cycle, improving F_LUC modeling, and informing effective climate mitigation policies.