Aversive learning is characterised by rapid learning which is highly resistant to extinction. This has likely developed as an evolutionary mechanism but it often becomes maladaptive. In this thesis I investigate computational mechanisms governing the learning of aversive beliefs, neural structures that are involved in the process and test whether pharmacological intervention can be used to amend learning. In a series of three experiments in healthy volunteers ranging in trait anxiety I employ a probabilistic aversive learning paradigm in behavioural (study 2 and 3) and neuroimaging setting (study 1). I first characterise the difference between acquisition and extinction, identifying that while acquisition learning is fast and reaches the target level, during extinction aversive expectations are consistently higher than true reinforcement rates. This extinction-specific overprediction increases over time but only in the low anxious group. High trait anxiety was associated with increased dissociability between acquisition and extinction which was driven by the activity in the dorsal anterior cingulate cortex. Computational modelling revealed that high anxiety is associated with the tendency to internally represent the learning environment as distinct states and switch between them. This finding was later supported by behavioural and computational evidence in study 2. The neuroimaging analysis suggests that state learning is processed by the dorsal anterior cingulate cortex, ventro-medial prefrontal cortex and the inferior parietal lobule. This finding highlights increased context-dependence of anxious individuals which bears importance to clinical practice. In study 3, I test the effect of Angiotensin-II receptor antagonist losartan on aversive learning showing that losartan specifically decreases learning rates without influencing perception or updating.