Various proposals that a stochastic event, "commitment," is the first and rate-limiting step in mammalian cell differentiation were tested in one cell type, B16C3 mouse melanoma cells. Differentiation (pigment production) was observed in time-lapse films and in cloned single cells. As predicted by all the theories, onset of differentiation was at widely variable times in different cells after stimulation; and selection experiments showed that little of the variability was genetic. Contrary to some theories, differentiation appeared unrelated to cell division. Two properties of the melanoma cells did not fit any of the theories: times of differentiation were highly correlated in sister cells; and differentiation could be reversed in a proportion of cells, which was highest at the lowest levels of pigmentation. Dedifferentiation was associated with cell proliferation, so that most pigmented clones were small and most unpigmented clones large. These findings are accommodated by a model in which functions associated with differentiation can switch on and off, but an inhibition of the off transition builds up in the on state.