<p>Artificial intelligence (AI) increasingly supports subsurface characterization by integrating heterogeneous geophysical, geochemical, and spatial data. However, existing studies remain fragmented across disciplines and often leave unresolved questions about cross-modal physical coupling, fusion design, uncertainty propagation, validation integrity, and decision relevance. This review synthesizes AI-based multimodal fusion methods for mineral exploration, geothermal systems, geological CO₂ storage, and radioactive-waste disposal. It organizes the field along orthogonal dimensions: the representation stage of fusion; the mechanism and strength of physical coupling; the training signal and model objective; the integration of prior knowledge; the output type; and the intended scope of generalization. The review critically examines scale and support mismatch, spatial autocorrelation, modality dominance, uncertain rock-physics and geochemical relationships, topological and conservation constraints in generative models, and validation leakage. It also distinguishes aleatoric, epistemic, methodological, and ontological uncertainty and considers how these uncertainties propagate through multimodal workflows. Emerging priorities include cross-modally coupled physics-informed learning, constrained generative simulation, foundation-model representations, and risk-aware monitoring. By clarifying conceptual distinctions and evidence limitations, the review provides a rigorous roadmap for developing interpretable, uncertainty-aware, and decision-relevant multimodal subsurface AI.</p>