Reducing Stigma In Digital Spaces Through Human-Centered AI - Substance use disorder (SUD) remains a major public health challenge in the United States. Despite substantial progress in scientifically supported treatment and recovery support, many individuals with SUD do not access or remain engaged in formal care. Persistent obstacles, including negative social judgment, cost, and structural constraints, limit help-seeking and recovery engagement. At the same time, recovery capital (RC)--the personal, social, and local resources that individuals draw upon to support long-term recovery--has been recognized as an important determinant of outcomes but is rarely examined in ways that capture its dynamic nature across recovery stages. As more individuals with SUD turn to online platforms such as Reddit to share their stories and seek support, we now have unprecedented access to rich, longitudinal narratives that can shed light on how stigma and RC are experienced in everyday life. Advances in natural language processing (NLP) and generative artificial intelligence (AI) present a powerful opportunity to analyze these narratives at scale and develop novel tools to support recovery. However, few studies have applied these methods to the SUD context or explored how insights from online recovery discourse might inform supportive interventions. This proposal aims to identify and characterize stigma and recovery capital in online SUD narratives using NLP methods and to develop a prototype generative storytelling tool to support recovery and self-affirming language. This early-stage work will lay the foundation for future intervention research and technology enabled recovery support. In Aim 1, we will analyze posts from drug-related and recovery-oriented subreddits using manual annotation and NLP pipelines. We will classify stigma types (internalized, anticipated, perceived) and identify sources of judgment (e.g., health care, family, institutions). Simultaneously, we will extract linguistic indicators of RC, including signs of resilience, social support, coping strategies, and access to local resources, and model how negative judgments and RC evolve over time and across user trajectories. In Aim 2, we will design and develop a proof-of concept generative storytelling prototype powered by generative AI models (e.g., GPT, Gemini, and Llama), grounded in the insights from Aim 1. The system will produce personalized recovery narratives that incorporate RC-enhancing elements and avoid language that reinforces self-blame and shame, helping users see themselves in more constructive roles. This project will advance our understanding of how stigma and recovery capital are expressed in real-world digital spaces and demonstrate the feasibility of a novel, scalable AI-driven approach to behavioral health support. By grounding the work in real-world narratives and leveraging the capabilities of generative AI, this work aligns with NIH priorities to support scientifically meritorious research that addresses urgent health needs.
