"We Are Tired of Explaining'": Designing LLM Roleplay Training for Community Health Workers in Rural India

Community health workers in the Global South are increasingly exposed to AI tools, yet the counseling work at the core of their role remains largely untouched by technology. We report a qualitative study with 20 Accredited Social Health Activists (ASHAs) in rural Rajasthan, India. Each session combined a standardized family-planning call with a beneficiary actor, a semi-structured interview, and an interaction with an LLM roleplay design probe. In the calls, participants often responded to social, material, or health concerns by returning to health-risk information, denying the concern, promising help without a plan, or naming contraceptive methods with limited explanation. A smaller set of responses engaged the concern, asked permission before involving family members, or left decisions with the caller. Probe reactions also exposed a deployment risk: some participants wanted to use generated audio as a persuasive voice during home visits or as a source of clinical information. We identify four design implications for AI-supported roleplay: make reflection rather than persuasion the training objective, keep rehearsal separate from authoritative or client-facing use, preserve situated judgment rather than supply scripted answers, and treat low-friction delivery as a deployment hypothesis to test. These findings identify concrete directions and boundaries for future AI-supported communication training with community health workers in the Global South.