Beyond transparency: why Traditional Chinese Medicine (TCM) need explainable artificial intelligence (XAI)

The integration of artificial intelligence (AI) with Traditional Chinese Medicine (TCM) is rapidly expanding, creating new opportunities for digital diagnosis, syndrome differentiation, prescription recommendation, multimodal clinical support, and knowledge mining. The central obstacle to this integration, however, lies beyond predictive performance and stems from a fundamental epistemological mismatch between data-driven AI and theory-driven TCM. Contemporary AI systems often operate through opaque statistical representations, whereas TCM depends on holistic, relational, and interpretive reasoning centered on syndrome differentiation and disease-mechanism inference. Consequently, explainable artificial intelligence (XAI) is required not merely to improve transparency but to serve as an epistemic interface that enables semantic translation between machine-discovered patterns and TCM clinical reasoning. This review argues that AI-TCM integration should be understood as an epistemological integration problem rather than a purely technical one. To address this problem, we introduce two conceptual lenses: the dual-layer opacity framework, which captures the superposition of algorithmic opacity and the theoretical opacity of TCM knowledge, and the semantic translation framework, which conceptualizes XAI as the process of mapping computational features and reasoning traces onto clinically meaningful and theory-consistent TCM concepts. On this basis, we critically synthesize major methodological pathways of TCM-XAI, including feature-attribution methods, visual explanation, intrinsically interpretable models, knowledge-guided reasoning, and large-language-model-based explanation infrastructures such as chain-of-thought and graph-based retrieval-augmented generation. Beyond methodological synthesis, we identify four core criteria for high-quality explanation in TCM: faithfulness, clinical relevance, theoretical coherence, and cultural integrity. We show that the principal challenges of TCM-XAI extend beyond accuracy and include annotation uncertainty, explanation validation, privacy-preserving interpretability, fairness across populations, and the risk of epistemic reduction or cultural misinterpretation. Finally, we outline a future research agenda centered on causal inference, neuro-symbolic reasoning, multimodal explanation benchmarks, clinician-centered evaluation, and regulatory standards for trustworthy TCM-AI systems. We argue that XAI should be understood not as a technical add-on but as the epistemic infrastructure required to build clinically trustworthy, culturally coherent, and scientifically robust AI ecosystems for Traditional Chinese Medicine.