Mental health problems have become a major global public health challenge, and traditional mental health service systems face persistent obstacles in early screening, assessment, and intervention, including limited resources, poor accessibility, and insufficient precision. Leveraging machine learning, deep learning, natural language processing, and multimodal data fusion, artificial intelligence (AI) is breaking service barriers, alleviating workforce shortages, and reducing stigma, thus enabling a continuous, dynamic, and personalized closed-loop paradigm for mental health management. This review synthesizes recent advances in AI-driven early screening through audio-video analysis, social media text mining, and wearable monitoring; early assessment through multimodal integration and biomarker modeling; and early intervention through personalized recommendation systems, digital therapeutics, and virtual agent-based interventions. Focusing on critical risks such as data privacy and security, algorithmic bias, algorithmic fairness, and ethical responsibility, the present paper further discusses future directions including constructing collaborative networks and data-sharing, developing brain-computer interfaces and neural regulation, and promoting global collaborative governance.