Abstract
Mental health disorders are highly prevalent worldwide, yet access to timely and effective mental health assessment (and care) remains limited. Artificial intelligence (AI) offers potential solutions, but the literature on its use in psychological assessment contexts has not been comprehensively mapped. This review aimed to systematise existing research, identify gaps, evaluate methodological limitations, and outline future directions. Using a librarian-approved search strategy, 7595 records were retrieved from major databases and screened independently by two coders. Following the eligibility assessment, 320 peer-reviewed articles were included. Studies showed wide variability in sample sizes (1-19,400,000) with no clear temporal trend. Most recruited clinical (21%) or general population (16%) samples from China (24%) or the United States (21%), and focused on depression (54%), anxiety (14%), suicidality (12%) or stress (8%). Supervised (75%) and deep learning (47%) approaches predominated, often with multiple algorithms compared (77% of the studies). Validation commonly relied on cross-validation and convergence with screening instruments, with relatively little use of DSM or ICD diagnostic criteria (71% used neither). Area-Under-the-Receiver-Operating-Characteristics-Curve (AUC) was the most frequently used performance metric, and unsupervised models achieved the highest average AUC. A marginal improvement in performance was evident from 2014 to 2025. Overall, AI shows promise as a psychological assessment tool, but progress is constrained by limited transparency, heavy reliance on self-report data, inconsistent use of validated diagnostic standards, a narrow focus on outcomes, and insufficient demographic and cultural analyses. Future research should prioritise interpretability, ethical and cultural responsiveness, multi-modal data, diverse samples, and clinically meaningful validation.