The human transcriptome contains millions of A-to-I editing sites arising from an unclear number of poorly characterized dsRNAs. Editing sites are often used to infer presence of dsRNA, but this method is limited by transcription levels, read depth, ADAR expression and cannot identify unedited dsRNA. To address these limitations, we developed dsRNAscan. Applying dsRNAscan to the human genome predicted 5 million dsRNAs. Genomic distribution of dsRNAs encompassing repetitive elements was widespread, but non-repetitive dsRNAs were sparse and enriched at chromosomal tips. dsRNAscan predicted hundreds of long, highly paired dsRNAs suspected to be immunogenic, but only one was in a 3'UTR, and thus likely to challenge cytoplasmic immune sensors. We observed several thousand editing enriched regions suspected to arise from intermolecular structures, and dozens of neuronally enriched dsRNAs conserved across vertebrates. This study offers the first comprehensive set of dsRNA annotations for the human genome, available as a resource at https://dsrna.chpc.utah.edu/.