Researchers at the University of Pennsylvania´s Leonard Davis Institute have built an automated system that uses Artificial Intelligence to assemble evidence-based, real-time social media campaigns aimed at hard-to-reach groups. Led by LDI senior fellow Dolores Albarracín and Man-pui Sally Chan, the team trained models on a dataset of HIV-related tweets and posts drawn from keyword searches and expert accounts, including federal agencies, nonprofits, and researchers. The system ranks messages for relevance and actionability and then recommends them to local health agencies for posting.
In a field test across agencies in 42 U.S. counties, the messages recommended by the system were posted six times more often than messages selected by a conventional approach, according to a June 17, 2025 article in PNAS Nexus. The project targeted men having sex with men, a population central to federal HIV elimination goals, and measured both participant ratings of persuasiveness and agency behavior. Albarracín called the method a ´game changer´ because it produces a ´living´ campaign that updates continuously, increasing message volume, timeliness, and community relevance.
The team describes a replicable pipeline: continuous access to social media streams, topic filtering, use of Artificial Intelligence to identify acceptable and actionable messages based on human-ground-truth data, recommendation to agencies, local vetting and adaptation, posting by agencies, and community reception and feedback. The researchers made the pipeline publicly available and say other organizations can implement it by gathering local human data and training models. They have not provided a formal cost estimate but suggested the approach is likely inexpensive compared with ad hoc message creation.
The authors acknowledge the risk that similar techniques could be misused to amplify misinformation and propose guardrails including platform moderation, audience migration away from misinfo-prone spaces, and frequent science-based messaging from public agencies. Next steps include a randomized controlled trial in Philadelphia to test effects on residents and measures beyond agency behavior. The study and its supplemental materials list numerous collaborators and provide the methodological backbone for scaling real-time, evidence-based public health messaging on social media.
