Maternal-effect selfish genes such as Medea or Scat act to kill progeny that do not bear a copy of the selfish gene present in the mother. Previous models of this system allowed for two types of allele, the selfish (killer) type and the sensitive (susceptible) wild-type. These models predict that the invasion conditions of the selfish allele are quite broad and that if invasion is possible a high frequency equilibrium is to be expected. The selfish element is therefore predicted to persist. Here a hypothetical third allele that neither kills nor is killed (i.e. insensitive) is considered. Such an allele could enter a population by recombination, mutation or migration. The incorporation of this third allele profoundly affects the dynamics of the system and, under some parameter values, it is possible for the spread of the insensitive allele to lead, eventually, to the fixation of the wild-type allele (reversible evolution). This is most likely if the death of progeny provides no direct benefit to the surviving sibs (i.e. in the absence of fitness compensation), as in insects without gregarious broods. Under these circumstances the selfish element cannot spread when infinitely rare, only after having risen to some finite frequency. A fitness cost to bearing the killer allele then causes its loss. However, if fitness compensation is found (e.g. in placental mammals) the invasion of the selfish element from an infinitely low level is possible for a wide range of costs and both stable coexistences of all three alleles and limit cycles of all three are then found. It is therefore to be expected that in mammals selfish maternal-effect genes are more likely both to spread and to persist than in insects, due to their different levels of fitness compensation.