Modern gene therapies aim to prevent the inheritance of mutant mitochondrial DNA (mtDNA) from mother to offspring by using a thirdparty mtDNA background. Technological limitations mean that these therapies may result in a small amount of maternal mtDNA admixed with a majority of thirdparty mtDNA. This situation is unstable if the mother's mtDNA experiences a proliferative advantage over the thirdparty mtDNA, in which case the efficacy of the therapy may be undermined. Animal models suggest that the likelihood of such a proliferative advantage increases with increasing genetic distance between mother and thirdparty mtDNA, but in real therapeutic contexts the genetic distance, and so the importance of this effect, remains unclear. Here we harness a large volume of available human mtDNA data to model random sampling of mother and thirdparty mtDNAs from real human populations. We show that even within the same haplogroup, genetic differences around 2080 SNPs are common between mtDNAs. These values are sufficient to lead to substantial segregation in murine models, over an organismal lifetime, even given low starting heteroplasmy, inducing increases from 5% to 35% over one year. Randomly pairing mothers and thirdparty women in clinical contexts thus runs the risk that substantial mtDNA segregation will compromise the beneficial effects of the therapy. We suggest that choices of "mtDNA donors" be based on recent shared maternal ancestry, or, preferentially, explicit haplotype matching, in order to reduce the potential for problems in the implementation of these therapies.