BACKGROUND: Verbal autopsy could be more widely used if interpretation by computer algorithm could be relied on. We assessed how InterVA-4 results compared with clinician review in diagnosing HIV/AIDS-related deaths over the period of antiretroviral (ART) roll-out. DESIGN: In the Karonga Prevention Study demographic surveillance site in northern Malawi, all deaths are followed by verbal autopsy using a semi-structured questionnaire. Cause of death is assigned by two clinicians with a third as a tie-breaker. The clinician review diagnosis was compared with the InterVA diagnosis using the same questionnaire data, including all adult deaths from late 2002 to 2012. For both methods data on HIV status were used. ART was first available in the district from 2005, and within the demographic surveillance area from 2006. RESULTS: There were 1,637 adult deaths, with verbal autopsy data for 1,615. Adult mortality and the proportion of deaths attributable to HIV/AIDS fell dramatically following ART introduction, but for each year the proportion attributed to HIV/AIDS by InterVA was lower than that attributed by clinician review. This was partly explained by the handling of TB cases. Using clinician review as the best available 'gold standard', for those aged 15-59, the sensitivity of InterVA for HIV/AIDS deaths was 59% and specificity 88%. Grouping HIV/AIDS/TB sensitivity was 78% and specificity 83%. Sensitivity was lower after widespread ART use. CONCLUSIONS: InterVA underestimates the proportion of deaths due to HIV/AIDS. Accepting that it is unrealistic to try and differentiate TB and AIDS deaths would improve the estimates. Caution is needed in interpreting trends in causes of death as ART use may affect the performance of the algorithm.