The Allostatic Model proposes that Alcohol Use Disorder (AUD) is associated with a transition in the motivational structure of alcohol drinking: from positive reinforcement in early-stage drinking to negative reinforcement in late-stage dependence. However, direct empirical support for this preclinical model from human experiments is limited. This study tests predictions derived from the Allostatic Model in humans. Specifically, this study tested whether alcohol use severity (1) independently predicts subjective responses to alcohol (SR; comprised of stimulation/hedonia, negative affect, sedation and craving domains), and alcohol self-administration and 2) moderates associations between domains of SR and alcohol self-administration. Heavy drinking participants ranging in severity of alcohol use and problems (N = 67) completed an intravenous alcohol administration paradigm combining an alcohol challenge (target BrAC = 60 mg%), with progressive ratio self-administration. Alcohol use severity was associated with greater baseline negative affect, sedation, and craving but did not predict changes in any SR domain during the alcohol challenge. Alcohol use severity also predicted greater self-administration. Craving during the alcohol challenge strongly predicted self-administration and sedation predicted lower self-administration. Neither stimulation, nor negative affect predicted self-administration. This study represents a novel approach to translating preclinical neuroscientific theories to the human laboratory. As expected, craving predicted self-administration and sedation was protective. Contrary to the predictions of the Allostatic Model, however, these results were inconsistent with a transition from positively to negatively reinforced alcohol consumption in severe AUD. Future studies that assess negative reinforcement in the context of an acute stressor are warranted.