<p>People differ in their tendency to experience negative emotions. This variability is largely captured by broad psychological constructs like neuroticism, whose facets include anxiety, depression, and stress vulnerability, among others. The amygdala and salience network have been assumed to underlie such dispositions, despite inconsistent evidence. We preregistered a comprehensive test of these and other competing hypotheses—accompanied by theory-agnostic machine learning prediction—using neural responses in the two most common emotional neuroimaging tasks (scenes and faces; <i>N</i> = 338/424). Evidence including Bayes factors indicated that neuroticism is not associated with any region, network, affective signature, or machine learning pattern, including the amygdala. Still, a brain-wide machine learning pattern robustly predicted the neuroticism facet <i>stress vulnerability</i> (r = .21), replicated in an independent dataset (r = .19). Predictive performance most strongly depended on somatomotor and visual networks, rather than salience network regions, relating stress vulnerability to cortical perception–action systems. Together with a multiverse analysis spanning 14 trait constructs and 1,176 models, our findings demonstrate the highly selective predictability of emotional dispositions from brain responses to common affective tasks. Therein, they highlight the importance of construct and task selection, while challenging the role of the most commonly used neural markers, including responses of the amygdala and salience network.</p>