•  
  •  
 

Abstract

This study introduces a generative artificial intelligence (GenAI) agent designed to autonomously evaluate, optimize, and personalize drug prevention campaigns across Facebook, Reddit, Instagram, and Twitter (X) using a 45,000-post multi-platform awareness corpus. Five state-of-the-art large language models, GPT-5-mini, Claude 3.5, Gemini 2.0, Qwen 2.5, and Mixtral, were examined under six structured prompting families, including Original, Role-based, Two-Stage Explicit Sensemaking, and Case-Based Adaptive Reasoning in short and long variants. Model outputs were assessed using a tri-metric framework comprising the Educational Rate (Edu-R), Violation Rate (Vio-R), and Misleading Awareness Score (MAS), supported by classical discrimination and agreement measures, including ROC-AUC and Cohen's Kappa, as well as influence-spread simulation. Results demonstrate that GPT-5-mini exhibits the strongest overall performance, achieving 95.10% accuracy, 96.22% precision, 94.55% recall, and 95.44% F1 score. Structured prompting substantially improved alignment and safety across all models, increasing GPT-5-mini's Edu-R from 78.12% under minimal instructions to over 95% under agent-based prompting. The Vio-Rs were reduced to low single-digit values, corresponding to approximately 96%-99% safety-aligned outputs. Influence-spread simulations further showed that cognitively rich prompts significantly enhance message diffusion, particularly in demographic clusters. The proposed GenAI agent establishes a scalable, evidence-driven foundation for real-time evaluation and personalization of drug prevention campaigns.

Download

Share

COinS