The Amazon is depended on a source of water for much of South America and its evapotranspiration forms a major component of the regional hydrological cycle. Under anthropogenic climate change, drying is projected over the Amazon by most global climate models (GCMs), including a devastating intensification of drought. Due to its unique hydrological cycle, we decompose the atmospheric moisture budget conditioned on drought over the Amazon across CMIP6, both historically and at end-of-century. Following a previous study, we define "feedback droughts" as multi-year meteorological droughts in which the deficit of evapotranspiration is greater than that of atmospheric moisture flux convergence (MFC). These droughts represent cases where an aridification of the surface exacerbates a precipitation deficit initially forced by atmospheric processes. Based on historical precipitation thresholds, CMIP6 models project between a 10% decrease and a 200% increase of drought per degree warming in the SSP3-7.0 scenario, with smaller (larger) increases under lower (higher) emissions scenarios. These increases are almost purely due to feedback droughts. Therefore, differences in representation of land—atmosphere feedbacks across GCMs lead to major disagreements in future drought projections over the Amazon. However, a secondary contribution to uncertainty arises from the initial MFC deficit, which we relate to disagreements in the pattern of sea-surface temperature (SST) change over the equatorial Pacific and Atlantic. Further, we evaluate the evolution of feedback droughts in the historical climate, comparing to reanalyses. Those in which the initial precipitation deficit is realistically exacerbated by land—atmospheric processes are almost all at the high end of the CMIP6 spread of future drought increases. And almost all such GCMs include a dynamic vegetation model. Thus, future drought increases over the Amazon are likely to be greater than most GCMs project, but uncertainty remains due to both uncertainty in SST changes and structural differences across dynamic vegetation models. An improved representation of land—atmosphere feedbacks in GCMs, as well as a continued effort to predict changes to tropical SST patterns, will further improve and constrain projections of drought in this critical region.