Individuals differ genetically in their susceptibility to particular diseases and their response to drugs. However, personalized treatments are difficult to develop, because disease susceptibility and drug response generally have poorly characterized genetic architecture. It is thus tempting to use the ethnicity of patients to capture some of the variation in allele frequencies at the genes underlying a clinical trait. The success of such a strategy depends on whether human populations can be accurately classified into discrete genetic ethnic groups. Despite the heated discussions and controversies surrounding this issue, there has been essentially no attempt so far to quantify the relative power of ethnic groups and geography at predicting the proportion of shared alleles between human populations. Here, we present the first such quantification using a dataset of 51 populations typed at 377 microsatellite markers, and show that pair-wise geographic distances across landmasses constitute a far better predictor than ethnicity. Allele sharing between human populations decays smoothly with increasing physical distance at a worldwide scale. We discuss the relevance of those patterns for the expected distribution of variants of medical interest. The distribution patterns for genes coding for simple traits are expected to be highly heterogeneous, as most such genes experienced strong natural selection. However variants involved in complex traits are expected to behave essentially neutrally and we expect them to fit closely our predictions based on microsatellites. We conclude that the use of ethnicity alone will often be inadequate as a basis for medical treatment.