Online job advertisements serve as a scalable entry point for deceptive and exploitative recruitment for human trafficking. We propose a novel Large Language Model (LLM)-driven framework for early-stage detection of recruitment risk in online job advertisements. We define recruitment risk and measure it using fifteen binary signals that reflect economic misrepresentation, information transparency, recruitment pressure and channels, and mobility and dependency. We further introduce a composite risk metric that aggregates these signals into an interpretable score, that allows for prioritization.
This additive formulation provides transparency by explicitly linking model outputs to underlying evidence, allowing stakeholders to understand why a job advertisement is flagged. The annotated dataset is then used to train transformer-based models (Longformer), which capture long-range textual dependencies and achieve strong predictive performance (86% accuracy), with substantial gains in previously under-detected risk categories. The proposed framework is novel in its integration of LLM-based annotation with a transparent, multi-dimensional risk scoring system. By transforming unstructured job advertisement text into auditable and interpretable risk metrics, this work advances the use of LLMs as foundational tools for scalable, explainable risk detection in complex textual environments.
