Artificial intelligence (AI) models for drug-induced liver injury (DILI) address heterogeneous prediction tasks, ranging from classifying molecular hepatotoxicity potential to estimating individual patient risk. Their reported performance may depend on the provenance of training labels, validation design, population characteristics, applicability-domain coverage, and calibration. To evaluate the extent to which published AI models provide evidence of reliable human DILI prediction, with particular attention to independent external validation, applicability domains, calibration, and clinically interpretable performance. Published literature from 2018–2026 was examined across chemical, toxicogenomic, experimental, and clinical prediction settings. Model-development data, DILI definitions, validation approaches, reported discrimination, calibration, and intended clinical applications were designated as primary extraction domains. Methodological assessment was informed by PRISMA 2020, TRIPOD+AI, and prediction-model risk-of-bias guidance. The identified literature encompasses drug-level classification, laboratory and toxicogenomic prediction, patient-level risk stratification, and differential diagnosis. These tasks employ different reference standards and validation designs. Evidence from held-out chemical datasets and experimental biological systems cannot be interpreted as direct validation of individual patient outcomes. Clinical models provide more directly relevant evidence, although temporal and geographical validation assess different aspects of transportability. The reliability of AI-supported DILI prediction depends on the prediction target, independence of evaluation data, applicability to the intended population, and the relationship between statistical performance and clinical decisions. No general claim of reliable human prediction is justified from development discrimination alone.