Abstract
Background: Artificial intelligence (AI) and machine learning (ML) are increasingly used in environmental epidemiology to integrate satellite, geographic, sensor, health-record, biomonitoring, and multi-omics data. Existing reviews are organized mainly by technology or domain; less explicit is how validation design, exposure-error structure, and inferential target determine what reported results can support. This review asks (i) for which inferential targets AI/ML is supported by the evidence, (ii) how strongly reported performance depends on validation design, (iii) whether complex algorithms improve on conventional baselines, and (iv) how exposure-error structure affects downstream health estimates. Methods: This is a critical narrative review (not a systematic review) of records in PubMed, Scopus, and Web of Science through 31 August 2026, organized by inferential target. Representative primary studies were chosen by explicit criteria and compared qualitatively, without pooling or scoring. Findings of the comparison: In the studies compared, reported accuracy depended strongly on validation design; in some comparisons, complex algorithms added little over simpler baselines; and the type and structure of exposure error, rather than prediction accuracy alone, determined whether health estimates were biased or merely imprecise. Interpretation and recommendations: General methodological cautions drawn from the wider literature (leakage, weak calibration, limited external validation, unrepresentative data, lack of transparency, privacy risks, and unequal accuracy across populations) are distinct from these comparative findings. Priorities include spatially and temporally blocked validation, conventional baselines, propagation of exposure-model error, subgroup auditing, and transparent documentation. Large language models may assist literature screening but require verification. AI should strengthen, not replace, epidemiological design, causal reasoning, and expert judgment.