INTRODUCTION: Verbal autopsy (VA) can be integrated into civil registration and vital statistics systems, but its accuracy in determining HIV-associated causes of death (CoD) is uncertain. We assessed the sensitivity and specificity of VA questions in determining HIV status and antiretroviral therapy (ART) initiation and compared HIV-associated mortality fractions assigned by different VA interpretation methods. METHODS: Using the WHO 2012 instrument with added ART questions, VA was conducted for deaths among adults with known HIV status (356 HIV positive and 103 HIV negative) in South Africa. CoD were assigned using physician-certified VA (PCVA) and computer-coded VA (CCVA) methods and compared with documented HIV status. RESULTS: The sensitivity of VA questions in detecting HIV status and ART initiation was 84.3% (95% CI 80 to 88) and 91.0% (95% CI 86 to 95); 283/356 (79.5%) HIV-positive individuals were assigned HIV-associated CoD by PCVA, 166 (46.6%) by InterVA-4.03, 201 (56.5%) by InterVA-5, and 80 (22.5%) and 289 (81.2%) by SmartVA-Analyze V.1.1.1 and V.1.2.1. Agreement between PCVA and older CCVA methods was poor (chance-corrected concordance [CCC] <0; cause-specific mortality fraction [CSMF] accuracy </=56%) but better between PCVA and updated methods (CCC 0.21-0.75; CSMF accuracy 65%-98%). All methods were specific (specificity 87% to 96%) in assigning HIV-associated CoD. CONCLUSION: All CCVA interpretation methods underestimated the HIV-associated mortality fraction compared with PCVA; InterVA-5 and SmartVA-Analyze V.1.2.1 performed better than earlier versions. Changes to VA methods and classification systems are needed to track progress towards targets for reducing HIV-associated mortality.