Social scientists want to discover how the features of high dimensional interventions---such as messages, images, and videos---influence outcomes. We introduce a new experimental design and statistical method that demonstrates how machine learning methods can be unified with causal inference techniques to both discover interventions of interest and credibly estimate their effect. We first prove that existing techniques are ill-equipped for discovering and credibly estimating causal effects. We then prove conditions that identify the effects of high dimensional interventions and introduce new machine learning models to uncover their effect. We apply our procedure to an intervention designed to assess the characteristics of candidates voters prefer and observational data describing how an agency responds to the public.