There is a growing interest in the study of human endogenous retroviruses (HERVs) given the substantial body of evidence that implicates them in many human diseases. Although their genomic characterization presents numerous technical challenges, next-generation sequencing (NGS) has shown potential to detect HERV insertions and their polymorphisms in humans. Currently, a number of computational tools to detect them in short-read NGS data exist. In order to design optimal analysis pipelines, an independent evaluation of the available tools is required. We evaluated the performance of a set of such tools using a variety of experimental designs and datasets. These included 50 human short-read whole-genome sequencing samples, matching long and short-read sequencing data, and simulated short-read NGS data. Our results highlight a great performance variability of the tools across the datasets and suggest that different tools might be suitable for different study designs. However, specialized tools designed to detect exclusively human endogenous retroviruses consistently outperformed generalist tools that detect a wider range of transposable elements. We suggest that, if sufficient computing resources are available, using multiple HERV detection tools to obtain a consensus set of insertion loci may be ideal. Furthermore, given that the false positive discovery rate of the tools varied between 8% and 55% across tools and datasets, we recommend the wet lab validation of predicted insertions if DNA samples are available.