Soil moisture (SM) and vapor pressure deficit (VPD) are key environmental drivers that regulate vegetation growth and thereby play a critical role in mediating the terrestrial carbon cycle. However, temporal inconsistencies in satellite derived vegetation indices have left long-term changes in vegetation sensitivity to SM and VPD under active debate. Here we estimated global vegetation sensitivity to SM and VPD during 1982-2021 by applying explainable machine learning with a temporal consistency prioritized normalized difference vegetation index (NDVI) dataset and hydro-climate anomalies. We reveal that SM generally enhances NDVI dynamics across most vegetated areas, particularly in arid and semi-arid regions, with its positive sensitivity increasing with aridity. Conversely, negative SM sensitivity is observed in boreal and humid areas due to confounding energy-related effects. VPD sensitivity generally shows a negative relationship with NDVI, except for an anomalous positive response on the Tibetan Plateau. Our analysis of inter-annual trends indicates increasing SM sensitivity in many regions (e.g., Horn of Africa, eastern North America), suggesting heightened drought risk, while VPD sensitivity generally decreased in southeastern Europe. State-of-the-art Earth system models capture the overall increase in global vegetation sensitivity to SM but do not reproduce the trend of vegetation response to VPD, possibly due to inability to accurately simulate the stomatal closure response of vegetation to continuously rising VPD. These findings highlight the critical importance of accurately simulating vegetation responses to water availability, as vegetation vulnerability to water stress is likely to intensify under ongoing climate change.