Digital Isolation by Design: Machine Learning Evidence of Psychological Harm from AI-Driven Social Media

Background Over 4.5 billion users worldwide experience algorithmically curated content, yet systematic evidence of psychological impacts remains fragmented. This creates urgent public health and policy challenges. Objective To quantify AI-driven recommendation algorithms' effects on mental health through an innovative meta-analysis integrating deep learning classification and graph neural network analysis. Methods We synthesized 30 studies (N = 47,892) examining anxiety, depression, loneliness, political polarization, and self-esteem. A CNN-LSTM model (87.3% accuracy) classified 50,000 + social media posts to identify vulnerability profiles, while graph convolutional networks mapped research knowledge structures. Results Random-effects models revealed significant adverse effects: anxiety (d = 0.42), depression (d = 0.38), loneliness (d = 0.51, largest effect), political polarization (d = 0.35), and self-esteem (d=-0.33). Adolescents on image-based platforms showed 57–71% larger effects. Deep learning identified three risk profiles, with high-risk users (19.6%) exhibiting clinically significant depression (PHQ-9 = 16.8). Passive consumption amplified loneliness (d = 0.52), while active engagement showed protective effects (d=-0.16). Conclusions Algorithmic content curation exerts meaningful psychological harms, particularly among vulnerable populations. Findings support evidence-based regulation prioritizing well-being over engagement maximization and demonstrate how AI methods can illuminate AI's own societal consequences.