A Fuzzy Evaluation of Sentence Encoders on Grooming Risk Classification

AI models struggle with coded and indirect language: The study found that fine-tuned AI models "fail to tag instances where the predator uses indirect speech pathways and coded language to evade detection," which is particularly problematic since predators deliberately use such language to avoid being caught.

Out-of-vocabulary words are a critical vulnerability: The research identified that misclassifications are "characterized by a higher presence of out-of-vocabulary (OOV) words in samples," meaning current AI systems don't adequately handle the informal, slang-heavy language typical of online grooming conversations.

Human judgment varies and matters for AI training: By using a "fuzzy-theoretic framework" to map assessments from law enforcement officers, real victims, and decoys, the paper demonstrates that different experts perceive grooming risk differently—suggesting AI systems need to account for this nuance rather than treating grooming detection as a simple yes/no classification.

Current AI approaches are insufficient for child safety: The study reveals that existing fine-tuned transformer models, despite previous research success, have significant blind spots in real-world grooming detection, indicating that deploying these systems without addressing their limitations could create a false sense of security.

Robustness improvements are essential before deployment: The authors explicitly call for "more robust models to identify coded language from noisy chat inputs," emphasizing that current technology needs substantial improvement before it can reliably protect children from sophisticated predators.