We demonstrated previously that 75% of infertile men with round, acrosomeless spermatozoa (globozoospermia) had a homozygous 200-Kb deletion removing the totality of DPY19L2. We showed that this deletion occurred by Non-Allelic Homologous Recombination (NAHR) between two homologous 28-Kb Low Copy Repeats (LCRs) located on each side of the gene. The accepted NAHR model predicts that inter-chromatid and inter-chromosome NAHR create a deleted and a duplicated recombined allele, while intra-chromatid events only generate deletions. Therefore more deletions are expected to be produced de novo. Surprisingly, array CGH data show that, in the general population, DPY19L2 duplicated alleles are approximately three times as frequent as deleted alleles. In order to shed light on this paradox, we developed a sperm-based assay to measure the de novo rates of deletions and duplications at this locus. As predicted by the NAHR model, we identified an excess of de novo deletions over duplications. We calculated that the excess of de novo deletion was compensated by evolutionary loss, whereas duplications, not subjected to selection, increased gradually. Purifying selection against sterile, homozygous deleted men may be sufficient for this compensation, but heterozygously deleted men might also suffer a small fitness penalty. The recombined alleles were sequenced to pinpoint the localisation of the breakpoints. We analysed a total of 15 homozygous deleted patients and 17 heterozygous individuals carrying either a deletion (n = 4) or a duplication (n = 13). All but two alleles fell within a 1.2-Kb region central to the 28-Kb LCR, indicating that >90% of the NAHR took place in that region. We showed that a PRDM9 13-mer recognition sequence is located right in the centre of that region. Our results therefore strengthen the link between this consensus sequence and the occurrence of NAHR.