Psychological correlates of dependence on generative and conversational AI: a systematic review and meta-analysis

Background. Dependence on generative and conversational artificial intelligence (AI) is measured with diverse instruments spanning addiction criteria, task reliance, and emotional bonding. No meta-analysis has pooled its psychological correlates across populations.

Methods. We conducted a pre-registered three-level meta-analysis (Zenodo: 10.5281/zenodo.23420672) searching eight electronic databases and citations of instrument papers on 16 October 2026, with citation chasing. Zero-order correlations were pooled using correlated and hierarchical effects models on Fisher's z with cluster-robust inference (CR2). Scale type (T1 addiction-type, T2 functional, T3 emotional, T4 mixed) was the primary moderator. Certainty was rated via GRADE.

Results. We included 127 samples (123 studies; N = 74,492; 29 countries; 70% university students; 39% mainland China). AI dependence correlated moderately with anxiety (r = 0.37, 95% CI [0.25, 0.48]), stress (r = 0.35 [0.25, 0.43]), depression (r = 0.32 [0.21, 0.42]), social anxiety (r = 0.36 [0.17, 0.53]), loneliness (r = 0.28 [0.18, 0.38]), and procrastination (r = 0.26 [0.12, 0.40]), and weakly and inversely with self-esteem (r = −0.18 [−0.35, −0.00]). All 95% prediction intervals crossed zero. Addiction-type and mixed scales showed no detectable difference; functional and emotional scales were rarely paired with clinical outcomes. Certainty was low to very low.

Conclusions. AI dependence moderately correlates with distress and self-regulation difficulties, but prediction intervals cross zero and evidence certainty is low. Functional and emotional dependence have rarely been studied alongside clinical outcomes.