Artificial intelligence has expanded drug repurposing from similarity-based candidate searches toward heterogeneous systems that combine molecular networks, transcriptomic states, genetics, biomedical knowledge graphs, clinical records, and other real-world data. The resulting literature is difficult to interpret because computational prediction, mechanistic plausibility, experimental confirmation, and patient-level validation represent different forms of evidence but are often discussed under the same language of “validation.” This scoping review maps how AI-enabled drug-repurposing studies construct and connect these evidence forms and examines where translation strengthens, remains conditional, or stops. A reproducible search and screening process identified 37 eligible evidence units from 2018–2026. Evidence was mapped by dominant data modality, computational method, repurposing objective, biomedical scale, and validation stage rather than pooled across incomparable performance measures. Knowledge-graph and network approaches formed a substantial part of the mapped literature, while transcriptomic and increasingly single-cell methods provided context-specific disease-state representations. Multimodal studies linked some of these molecular representations to genetics or patient data, whereas real-world evidence supported both hypothesis generation and post-nomination evaluation. Only a subset of computational programmes traversed more than one validation setting. The evidence therefore supports viewing AI drug repurposing as a set of heterogeneous translational pathways rather than a single prediction task. Progress depends less on maximizing an isolated model score than on establishing whether successive evidence sources provide sufficiently independent, biologically relevant, and clinically interpretable support for the proposed new use.