Large-scale high-throughput sequencing data sets have been transformative for informing clinical variant interpretation and for use as reference panels for statistical and population genetic efforts. Although such resources are often treated as ground truth, we find that in widely used reference data sets such as the Genome Aggregation Database (gnomAD), some variants pass gold-standard filters, yet are systematically different in their genotype calls across genotype discovery approaches. The inclusion of such discordant sites in study designs involving multiple genotype discovery strategies could bias results and lead to false-positive hits in association studies owing to technological artifacts rather than a true relationship to the phenotype. Here, we describe this phenomenon of discordant genotype calls across genotype discovery approaches, characterize the error mode of wrong calls, provide a list of discordant sites identified in gnomAD that should be treated with caution in analyses, and present a metric and machine learning classifier trained on gnomAD data to identify likely discordant variants in other data sets. We find that different genotype discovery approaches have different sets of variants at which this problem occurs, but there are characteristic variant features that can be used to predict discordant behavior. Discordant sites are largely shared across ancestry groups, although different populations are powered for the discovery of different variants. We find that the most common error mode is that of a variant being heterozygous for one approach and homozygous for the other, with heterozygous in the genomes and homozygous reference in the exomes making up the majority of miscalls.