Abstract
<p>Computational psychiatry has sought the mechanisms of anxiety in domain-general biases in learning and planning. We argue this is the wrong target. The small, inconsistent effects that characterize this literature are not a failure of computational methods but a symptom of measuring generic parameters when trait anxiety is generated by person-specific world models — the causal beliefs an individual holds about threat, control, and the self across life domains. We distinguish the task model (the structure of a given experiment) from the world model, and show that theory-consistent findings are precisely those conditioned on context and domain. Trait anxiety, on this view, forms through domain-specific learning over development and is best characterized not by how individuals learn within a novel task, but by how world models are deployed in relevant domains. We close by arguing that inferring these idiographic world models — drawing on developmental, phenomenological, and language-model methods — is the central task for the next generation of computational models of anxiety.</p>