Does AI predict personality traits? A meta-analytic review of machine-predicted Big Five scores: Convergence and caveats

To what extent do machine-learning (ML) models predict self-reported Big Five personality traits? We conducted two integrated three-level meta-analyses of 119 primary studies ( k = 48 correlation-based and k = 71 accuracy-based), examining both regression and classification prediction performance. Across the five traits, ML-predicted and self-reported scores showed moderate-to-strong convergence, with mean corrected correlations ranging from ρ = 0.40 to 0.47 (95% CIs = 0.32–0.55; 80% credibility intervals = 0.07–0.81). Classification models achieved mean accuracies ranging from 64% to 69% (95% CIs = 60%–72%; 80% credibility intervals = 47%–84%). Moderator analyses indicated that data modality, particularly multimodal input, and its interaction with algorithmic complexity explain a substantial proportion of the true heterogeneity. Using meta-analytic structural equation modeling, we further compared the latent structures of ML-predicted and self-reported Big Five traits. The strong factor congruence (φ = 0.94) indicated fair similarity between the two structures. Overall, ML models show capability in replicating self-reported personality scores, but their wide credibility intervals, variation across modalities and algorithms, as well as proximity to self-reports' latent structure warrant caution against using ML-based personality estimates for high-stakes applications. Our findings provide benchmarks for future model development and highlight the need for their further validation.