SAND: An AI-Integrated WhatsApp Chatbot Combining a Decision-Game Simulation with the Delphi Technique for Drug-Abuse Awareness Among Youth in Sudan

Abstract

Drug abuse is a socio-economic and public health crisis among youth in Sudan, intensified by armed conflict, economic instability, and the limitations of fear-based awareness campaigns. This study developed and evaluated SAND, an artificial intelligence (AI)-driven conversational agent deployed on WhatsApp to provide scalable, interactive, and culturally tailored prevention. SAND integrates a branching Decision-Game simulation with the Delphi technique to promote drug-abuse awareness, critical thinking, and safe decision-making among Sudanese youth aged 18–24. A multidisciplinary panel of 15 experts in addiction prevention, mental health, Islamic counseling, and artificial intelligence completed a three-round Delphi process to develop, validate, and refine culturally sensitive risk scenarios. These scenarios were delivered through WhatsApp, where participant decisions were logged, session-level risk scores were updated, immediate consequence-based feedback was provided, and high-risk users were triaged to confidential, human-reviewed support channels. The system incorporates a four-layer safety architecture comprising a Crisis/Safety filter, Intent-Classification router, Retrieval-Augmented Generation (RAG) knowledge engine, and Anonymized Logging/Escalation layer. Efficacy was assessed using a randomized controlled trial with a pretest-posttest control-group design involving 200 participants. The experimental group interacted with SAND, whereas the control group received non-interactive awareness content. Results showed statistically significant and greater improvements in the experimental group in drug-risk awareness, risk-vulnerability reduction, decision-making self-efficacy, and adaptive coping (all p < .001). These findings suggest that combining expert consensus with interactive behavioral learning can support privacy-preserving substance-abuse prevention in crisis-affected, low-resource settings. However, the unusually large effect sizes warrant cautious interpretation, and longer-term, multi-site replication is recommended before broad implementation.