Artificial Intelligence in Skin Cancer Detection: A Narrative Review of Diagnostic Evidence and Clinical Applications

Artificial intelligence has shown strong performance on selected skin lesion classification tasks, but accurate image classification alone does not establish clinical benefit. Differences in data provenance, disease spectrum, evaluation design and human interaction complicate interpretation of published results. The main objective of the study was to critically examine representative evidence on artificial intelligence for skin cancer recognition, with emphasis on dermoscopic imaging, public datasets, diagnostic evaluation and clinical translation. A focused narrative synthesis was developed using targeted literature searches and verification of primary publications, dataset descriptors and reporting guidelines through 21 October 2026. This was not a systematic review or meta-analysis. No new model was trained and no patient-level data were analysed. Early retrospective studies established the feasibility of convolutional neural networks for selected diagnostic comparisons. Subsequent research demonstrated both potential benefits of clinician–AI collaboration and susceptibility to misleading support. Prospective multicentre evaluation provides stronger evidence of diagnostic performance, while showing clinically relevant sensitivity–specificity trade-offs. Public datasets support reproducibility but differ in imaging modality, diagnostic labels, reference standards and sampling. Class imbalance, related images across data partitions, skin-tone disparities, acquisition artifacts and distribution shift remain central concerns. Explanations may influence decisions differently depending on user expertise and the order of presentation. The evidence supports continued evaluation of AI as assistance for defined clinical tasks. Responsible translation requires independent validation, transparent reference standards, subgroup analysis, calibrated operating thresholds, and prospective measurement of patient and service outcomes. Comparative evaluation of the complete clinician–AI workflow should take priority over isolated claims of architectural superiority.