Human immunodeficiency virus (HIV) infection progresses to AIDS following an asymptomatic period during which the virus is thought to evolve towards increased fitness and pathogenicity. We show mathematically that progression to the strongest HIV-induced pathology requires evolution of the virus towards reduced replicative fitness in vivo. This counter-intuitive outcome can happen if multiple viruses co-infect the same cell frequently, which has been shown to occur in recent experiments. According to our model, in the absence of frequent co-infection, the less fit AIDS-inducing strains might never emerge. The frequency of co-infection can correlate with virus load, which in turn is determined by immune responses. Thus, at the beginning of infection when immunity is strong and virus load is low, co-infection is rare and pathogenic virus variants with reduced replicative fitness go extinct. At later stages of infection when immunity is less efficient and virus load is higher, co-infection occurs more frequently and pathogenic virus variants with reduced replicative fitness can emerge, resulting in T-cell depletion. In support of these notions, recent data indicate that pathogenic simian immunodeficiency virus (SIV) strains occurring late in the infection are less fit in specific in vitro experiments than those isolated at earlier stages. If co-infection is blocked, the model predicts the absence of any disease even if virus loads are high. We hypothesize that non-pathogenic SIV infection within its natural hosts, which is characterized by the absence of disease even in the presence of high virus loads, could be explained by a reduced occurrence of co-infection in this system.