Background: This study aims to systematically analyze the structural transition of AI in mental health, differentiating between passive tools and autonomous agents, and to propose a governance framework to facilitate responsible integration or mitigate integration risks. Methods: Employing a semi-systematic approach, we screened records from IEEE Xplore, PubMed, and ACM DL, ultimately analyzing 61 included studies. We track the transition from the first paradigm, AI-as-Tool (AI-T) to the second paradigm, AI-as-Agent (AI-A). Results: Early empirical evidence suggests that AI-A systems may assist in fostering preliminary working alliances and demonstrate potential for symptom reduction in controlled settings; however, their efficacy cannot currently be equated with, nor serve as a replacement for, standard low-intensity clinical care. Conclusions: To mitigate these risks, we propose the Tiered Human–AI Healing Ecosystem (THHE) for mental health. This framework utilizes dynamic autonomy modulation—automatically restricting AI agency based on real-time risk markers—to manage transitions between AI-led support and human-led care, promoting clinical safety.
Learning potential assessment is important for understanding individual differences in learning processes and cognitive development. Traditional assessments mainly rely on questionnaires or behavioral data collected during tasks, emphasizing external performance while lacking sensitivity to learners’ internal cognitive activities. Therefore, they provide limited evidence about the neurocognitive mechanisms underlying individual learning potential. This study investigated whether task-evoked EEG responses acquired during a dynamic learning paradigm could characterize individual differences in task-specific learning performance under graduated instructional support. Participants were 50 undergraduate students who completed a dynamic cognitive task designed to evaluate learning performance under graduated instructional support. An electroencephalogram (EEG)-based Multi-scale Learning Potential Assessment (MS-LPA) framework was developed, including a Dynamic Learning Paradigm with progressively complex cognitive tasks and responsive cueing; multiscale Neural Complexity Representations quantifying signal complexity and inter-regional synchronization; and a ResNet-18 model with SHAP-based interpretability for classification and biomarker identification. Experiments showed that MS-LPA captured task-evoked neural responses and outperformed traditional outcome-driven evaluation schemes. Entropy-based features achieved an accuracy of 0.77, and functional connectivity matrices reached 0.83. These findings demonstrate the feasibility of using EEG-derived neural representations to differentiate individual differences in task-specific dynamic learning performance and provide insights into its underlying neural mechanisms.