The frictional strength of faults controls the stability and the dynamics of slip in diverse natural phenomena including earthquakes, induced seismicity, and landslides. It has been shown that geological faults and rock discontinuities in general are rough in a wide range of scales but the dependency of slip dynamics on surface roughness for a large spectrum of heights has never been measured. Here we show specifically how slip dynamics dramatically varies as a function of surface roughness using direct shear experiments performed on laboratory-generated faults in Diabase. We demonstrate experimentally that under relatively low normal stress of 5 MPa, stick-slip oscillations, commonly referred to as laboratory earthquakes, only occur in a very limited range of roughness values within which a specific level, defined here as the critical roughness, triggers the highest amplitude of oscillations. Sliding across roughness higher or lower than critical is typically stable. Using monitored vertical motions through slip (dilation) coupled with numerical modelling we show that sliding on extremely “smooth” surfaces is typically stable because the very small height of asperities does not allow for the nucleation and motion arrest required for ‘stick’ phases to ensue, whereas sliding on extremely rough “fractured” surfaces is typically characterized by shearing through the tall asperities. Sliding across “sawcut” surfaces, however, is found to be particularly susceptible to stick slip deformation, and the shear motion is shown to be purely dilatant where the dilation and compression of the sliding interface are in phase with the stick slip oscillations. We conclude therefore that dilatant shear across moderately rough interfaces is a prerequisite for stick slip oscillations and consequently, for sliding instability.