We study how voters process information through two experiments around Indian elections. In a large-scale experiment, we show that providing voters information about candidates’ criminal charges increases votes for clean candidates and reduces votes for criminal politicians, with larger penalties for candidates facing more and serious charges. A follow-up experiment replicates these results and identifies two mechanisms. First, information facilitates learning: voters form more accurate beliefs and evaluate criminal candidates less favorably. Second, using direct measures of voter attention, we show that information makes criminality more salient, and increases its weight in voting decisions. Salience effects are larger when information is surprising or highlights contrast, but do not vary with decision relevance, consistent with bottom-up attention. Causal forest estimates provide further evidence that learning and salience are both important drivers of changes in voting behavior. We develop a simple model that integrates salience theory into a standard probabilistic voting framework to explain our results.