The growing catalogue of structural variants in humans often overlooks inversions as one of the most difficult types of variation to study, even though they affect phenotypic traits in diverse organisms. Here, we have analysed in detail 90 inversions predicted from the comparison of two independently assembled human genomes: the reference genome (NCBI36/HG18) and HuRef. Surprisingly, we found that two thirds of these predictions (62) represent errors either in assembly comparison or in one of the assemblies, including 27 misassembled regions in HG18. Next, we validated 22 of the remaining 28 potential polymorphic inversions using different PCR techniques and characterized their breakpoints and ancestral state. In addition, we determined experimentally the derived allele frequency in Europeans for 17 inversions (DAF = 0.01–0.80), as well as the distribution in 14 worldwide populations for 12 of them based on the 1000 Genomes Project data. Among the validated inversions, nine have inverted repeats (IRs) at their breakpoints, and two show nucleotide variation patterns consistent with a recurrent origin. Conversely, inversions without IRs have a unique origin and almost all of them show deletions or insertions at the breakpoints in the derived allele mediated by microhomology sequences, which highlights the importance of mechanisms like FoSTeS/MMBIR in the generation of complex rearrangements in the human genome. Finally, we found several inversions located within genes and at least one candidate to be positively selected in Africa. Thus, our study emphasizes the importance of careful analysis and validation of large-scale genomic predictions to extract reliable biological conclusions.