Performance Gaps in AI Social Cognition: A Comprehensive Meta-Analysis of Current Limitations and Human-AI Collaborative Approaches

Background: Current artificial intelligence systems demonstrate significant per- formance gaps in social cognitive tasks compared to human capabilities, raising questions about fundamental limitations and optimal deployment strategies. Methods: We conducted a systematic meta-analysis of 42 peer-reviewed studies encompassing 11,989 human participants and 127 distinct AI systems across six so- cial cognitive domains. We analyzed publicly available benchmark datasets includ- ing EmotiNet, MELD, SocialIQA, CommonsenseQA, and comprehensive Roboflow computer vision datasets totaling over 40,000 images. Results: Meta-analysis reveals consistent performance gaps across all domains (Cohen’s d = 0.85–2.31, all p < 0.001). AI systems achieve 45–78% of human performance depending on task complexity. Our human-AI collaborative framework demonstrates 15–45% performance improvements over pure AI approaches. Conclusions: Current AI architectures face systematic challenges in social cogni- tive tasks, particularly those requiring complex contextual integration. Human-AI collaboration represents a promising approach for addressing these limitations in practical applications.