AI Optimizes HIV Peer Referrals in South Africa Field Trial

Harvard researchers, led by computer science professor Milind Tambe, are testing an artificial intelligence protocol in South Africa to improve peer referrals for HIV testing and treatment through the HIV-LIFT project. The initiative, led in part by Wits University principal investigator Alastair van Heerden and involving Harvard's Teamcore group, the World Health Organization and other partners, aims to reach groups such as men and adolescents who are less likely to seek testing, diagnosis and care. South Africa has about 8 million people living with HIV, the highest number globally, and men have historically faced lower uptake of services because of stigma, work schedules and related barriers. The field trial is underway in KwaZulu-Natal province and centers on Coach Mpilo, a free, confidential AI health coach delivered through WhatsApp. Designed for South African users, the chatbot provides information and support related to HIV testing, treatment and other health issues, and it has enrolled more than 150,000 users, only about 15 percent of whom are men. The app includes a financial incentive for users to refer friends, family members or contacts who may need HIV services. The central question for the study is how to allocate limited incentives most effectively when only a subset of users can be encouraged to recruit others. To address this, Harvard has added a machine learning and optimization approach that identifies existing users most likely to refer people in underserved groups. The model is intended to supplement, not replace, clinics, health workers and the social relationships that enable peer referrals. The team built the approach from earlier HIV-LIFT campaigns and adjusted its design after learning that complex simulations were difficult for field partners to interpret. Retrospective analysis of about 16,000 historical incentives over two months suggested the nudging method could generate more successful referrals than alternative strategies. The planned comparison will test two approaches: randomly selecting eligible users for outreach, and applying an AI-generated priority list to narrow the incentive-eligible pool. If the AI-guided targeting proves effective and transparent, the priority-group strategy could be adopted permanently in Coach Mpilo or expanded to other applications. Tambe, who has pursued AI for social good for more than two decades and has focused on HIV prevention since 2013, said the goal is to help people get HIV information, improve care access and support health systems in practical, real-world settings.