The Precision-Weighting Dysregulation Hypothesis: An Active Inference Framework for Criminal Behaviour, Differential Pathways, and Testable Predictions

Background: Criminal behaviour has been modelled through sociological, psychological, and biological lenses, yet mechanistic integration across subjective experience, neural processes, and social niche remains limited. Existing frameworks identify risk factors and correlates but rarely specify the computational mechanism through which adverse environments produce criminal inference. Framework: Active Inference (Free Energy Principle) provides a unified computational architecture in which perception, action, and learning are driven by the minimisation of variational free energy under precision-weighted prediction errors. We apply this framework to criminal behaviour, proposing that crime arises from systematic dysregulation of precision weighting within hierarchical generative models, shaped by the interaction between neural architecture and social niche. Model: We specify three distinct pathomechanisms: (1) niche-induced prior rigidity, mapping onto life-course-persistent delinquency; (2) vertical hierarchy collapse, mapping onto crimes of passion; and (3) failed inference repair, explaining recidivism as a systemic prediction of the model. Differential pathways: The model is transdiagnostic: the same upstream precision-weighting dysregulation can produce criminal behaviour, depression, anxiety, or somatic conditions depending on specific niche configuration, policy-precision parameters, and available action models. We formally specify the conditions under which the externalising (criminal) pathway dominates and argue, from a medical perspective, that this pathway warrants particular attention because it is therapeutically underserved, generates additional self-harm through moral injury, and is addressed by a justice system that systematically reinforces the underlying pathology. Predictions: Three falsifiable predictions are derived, each targeting a different pathway: (1) shifted dopaminergic value-learning in juvenile offenders, (2) PFC–amygdala decoupling in crimes of passion, and (3) superiority of inference-based rehabilitation over standard reintegration programmes. The proposed Inference Repair Programme (IRP) targets the generative model itself—not merely behavioural policies—and therefore aims at restoration of inferential flexibility rather than re-normalisation. Epistemic status: The Free Energy Principle as a mathematical principle is unfalsifiable, analogous to Hamilton’s principle of least action. Active Inference as a process theory is falsifiable. Empirical validation to date is in a stage of productive exploration: models explain behavioural data reasonably well but have been applied predominantly post hoc (Hodson et al., 2024). The present predictions offer an opportunity to test Active Inference in a novel domain with genuinely novel predictions.