AI Agents Model Psychological Disorders for Computational Psychiatry

This research introduces a novel approach to modeling psychological disorders within artificial intelligence, specifically reinforcement learning (RL) agents. Instead of hand-tuning individual disorders, the team developed a "dose-controllable" manipulation of cognitive appraisal signals within an appraisal-guided PPO agent. This allowed them to express seven distinct disorders—anxiety, mania, OCD, depression, impulsivity, addiction, and PTSD—each controlled by a single parameter grounded in computational psychiatry. The study involved over a thousand runs, demonstrating a graded, monotone dose-response for each disorder, which control groups could not replicate. Beyond the induced effects, the research uncovered several emergent findings: disorders self-organized into a two-dimensional affective space where mania mirrored anxiety; removing a disorder knob remitted reward distortion disorders but not avoidance disorders, which instead recovered with graded exposure; and simultaneous disorder manipulations showed non-additive interactions, yielding testable comorbidity predictions. Crucially, the same "knobs" used to induce a disorder could also model its treatment. The framework's generalizability was further shown by transferring three disorder knobs (depression, addiction, anxiety) to a 3D pixel environment using a standard convolutional agent, confirming cross-assay dissociation across both domains. This indicates the framework is not limited to grid worlds or PPO's appraisal critic.