An AI-Assisted Tailoring of Actionable HIV/HCV and Substance Use Treatment Messages from Social Media among People Who Use Drugs in Philadelphia

Background: Social media contains abundant HIV-related content, but most posts are not directly actionable and rarely connect audiences to local prevention, care, or substance use treatment services. We adapted the ARMT-HIV tool to develop an AI-assisted tailoring approach to (1) identify actionable HIV-related messages and (2) pair messages with location-specific HIV/HCV prevention and care resources and substance use treatment information. Methods: We collected, characterized, and filtered HIV/HCV/substance use tweets (including expert-account sources) from a pool of 107,000 X messages, tagged messages by prevention/care continuum stage and demographic targeting using keywords, and paired messages with a locally curated service database (Philadelphia) developed with a Community Advisory Board. We conducted a within-subject online message evaluation experiment among N=100 people who use drugs recruited via the Public Health Management Corporation (PHMC) in Northeast Philadelphia (45% female; 55% male; 64% Black; 24% White). Based on participant-reported HIV/HCV status, gender, and race/ethnicity, each participant viewed 24 messages (12 adapted ARMT-HIV “experimental” vs 12 control, randomized order) and rated outcomes including actionability, appropriateness, acceptability, sharing intentions, persuasiveness, perceived effectiveness, personal relevance, and intention to change behaviors. Analyses used mean comparisons and multilevel models clustered by participant. Results: Across outcomes, adapted ARMT-HIV experimental messages were rated more favorably than control messages (all ps ≤ .040 in mean comparisons). For example, actionability was higher for experimental vs control messages (M=3.11 vs 2.82, 4-point scale; t=5.42, p<.001), as were appropriateness (2.81 vs 2.64; t=4.06, p<.001), acceptability (2.99 vs 2.85; t=3.17, p=.002), and sharing intentions (2.97 vs 2.80; t=4.10, p<.001). Multilevel models showed a significant positive effect of tailoring across all outcomes (e.g., b=0.28 for actionability; b=0.23 for perceived effectiveness; all p<.001). Conclusions: An AI-assisted pipeline combining actionability classification, rule-based tagging, and pairing with local HIV/HCV and substance use services produced messages that PWUD rated as more actionable, relevant, and persuasive than control messages. These findings support further effectiveness and implementation studies to test real-world engagement and behavioral outcomes.