Highly multiplex DNA sequencers have greatly expanded our ability to survey human genomes for previously unknown single nucleotide polymorphisms (SNPs). However, sequencing and mapping errors, though rare, contribute substantially to the number of false discoveries in current SNP callers. We demonstrate that we can significantly reduce the number of false positive SNP calls by pooling information across samples. Although many studies prepare and sequence multiple samples with the same protocol, most existing SNP callers ignore cross-sample information. In contrast, we propose an empirical Bayes method that uses cross-sample information to learn the error properties of the data. This error information lets us call SNPs with a lower false discovery rate than existing methods.