Scholars have suggested that White American support for welfare is related to beliefs about the racial composition of welfare recipients. While a host of observational studies lend credence to this view, it has not yet been tested using the tools of randomized inference. In this study, we do this by conducting two incentive-compatible experiments (n = 9,775) in which different participants are randomly given different signals about the share of welfare recipients who identify as Black and White. Our analysis yields four main findings. First, 86% of respondents greatly overestimate the share of welfare recipients who are Black, with the average respondent overestimating this by almost a factor of two. Second, White support for welfare is inversely related to the proportion of welfare recipients who are Black―a causal claim that we establish using treatment assignment as an instrument for beliefs about the racial composition of welfare recipients. Third, just making White participants think about the racial composition of welfare recipients reduces their support for welfare. Fourth, providing White respondents with accurate information about the racial composition of welfare recipients (relative to not receiving any information) does not significantly influence their support for welfare.