Experts must often choose whether to communicate using numbers or language when sharing information with policymakers and the general public. While there are many reasons why experts would send messages using language or numbers, our study emphasizes the role of message precision: numbers represent precise estimates, whereas language is imprecise because one word can be interpreted as many numerical values. We run a large-scale experiment in which we vary message senders’ incentives to directionally persuade their audience or to convey accurate information. We look at choices between message formats, finding that directional incentives increase the likelihood of using language to communicate. We also analyze preferences for messages within each format, finding distortions for both numerical and language messages. Message senders are more effective at persuading when they use language, even when message receivers know their incentives. All effects are similar when senders use imprecise probability intervals instead of language, suggesting that the imprecision of language explains the patterns that we observe. We discuss an application involving the communication of research evidence to policymakers.