Harvard researchers compare AI targeting with random outreach for HIV referrals

Limited money for HIV outreach creates a consequential choice: which people should receive incentives to bring friends into care? Researchers in South Africa are testing whether AI can improve that choice. A field trial in KwaZulu-Natal compares AI-guided targeting with random outreach through Coach Mpilo, an existing WhatsApp health coach, to encourage peer referrals for HIV services.

Harvard’s engineering school described the ongoing trial on October 20. The team, led by computer scientist Milind Tambe, is adding a machine-learning approach to an existing referral program rather than introducing a new chatbot. The aim is to identify users most likely to recruit people from underserved groups, including men and adolescents, for HIV services.

Coach Mpilo is a free, confidential AI health coach available through WhatsApp, designed for South Africans seeking information and support about HIV and other health issues. Harvard says it has more than 150,000 enrolled users, but only about 15% are men. Men have historically been less likely to seek HIV testing and treatment, with stigma and long working hours among the barriers.

Choosing whom to encourage

The referral program offers existing users financial incentives to recruit friends, relatives and other contacts who might need HIV services. Its budget can support encouraging only a limited number of people. The researchers’ task is therefore narrower than deciding who needs care: they are prioritizing which existing users to approach about bringing others into the program.

The AI combines machine learning with optimization—a method for choosing among options within constraints—to narrow the incentive-eligible pool. It uses data from earlier incentive campaigns to produce a fixed priority list before outreach. That design makes the field comparison a test of a prepared targeting policy, rather than a system continually changing its priorities during the campaign.

Historical promise meets practical limits

The reason to test this approach comes from a retrospective evaluation of approximately 16,000 historical user nudges collected over two months. Researchers estimated that AI-guided targeting would generate more successful referrals than the other strategies they evaluated. That is an estimate based on past campaigns, not a measured improvement from the current field trial.

The team also changed course on complexity. Graduate student Akseli Kangaslahti said an earlier, more complex simulation algorithm produced results that regular users could not easily interpret. The researchers learned from that work, he said, even though they could not directly use the algorithms they had proposed.

Transparency became a requirement for the prioritization approach. Tambe said local partners proposed policy restrictions intended to ensure fairness and equal treatment for users with identical profiles. Those restrictions sit alongside the goal of improving referrals: the method also has to remain simple enough for people running the program to understand and practical enough to fit its existing systems.

The trial belongs to HIV-LIFT, a Gates Foundation-funded, multi-year initiative led by Wits University’s Alastair van Heerden. Partners include Harvard’s Teamcore group and the World Health Organization. The researchers will use the trial to determine whether the priority-group approach could become a permanent part of Coach Mpilo or other applications. It is intended to support—not replace—clinics, health workers and the personal relationships behind peer referrals.