Learning to Detect Recruitment Fraud: A Critical Review of Dataset Lineage, Evaluation, and Decision Support

Preprint — critical narrative review; not peer reviewed. Machine learning studies report strong performance on fraudulent job advertisements, but benchmark scores alone provide limited evidence of reliable recruitment screening. This critical narrative review examines 13 purposively selected empirical studies published between 2017 and 2026, alongside dataset documentation and methodological literature. Ten core studies have explicitly identified ancestry in the Employment Scam Aegean Dataset; corpus cleaning, translation, and renaming therefore require provenance checks before being interpreted as independent validation. Reported results also differ in test populations, resampling boundaries, and metric averaging. A within-study resampling comparison illustrates how protocol changes alter the interpretation of accuracy, while recent cascades show that threshold adjustment and routing errors must be separated from the contribution of a second model. Contextual and human-input approaches broaden the available evidence, but subjective trust and predictive accuracy answer different questions. The review provides a study-level protocol audit and a reporting framework covering split independence, fraud-class performance, calibration, explanations, and operational outcomes. Priorities include independently collected multilingual advertisements, threshold-matched comparisons, and prospective assessment of human decisions. The resulting synthesis supports more informative evaluation rather than a ranking of architectures across incompatible experiments. Supporting materials: study-level evidence ledger, protocol audit, reference-verification record, search and evidence log, and LaTeX manuscript source. The core evidence includes 10 full-text appraisals and 3 appraisals limited to abstracts and publication records. AI assistance disclosure: This review develops an earlier university-project manuscript supplied by the author. In October 2026, OpenAI Codex assisted with literature discovery, source checking, evidence extraction and synthesis, drafting, editing, and LaTeX/figure preparation. Anthropic Claude provided critical feedback. Exact model versions were not recorded. The author is responsible for the final manuscript and its interpretation. Former affiliation: Department of Computer Science, Namal University, Mianwali, Pakistan.Corresponding author: Jalal Ur Rahman; [email protected]: none. Conflicts of interest: none declared.