Frictional forces were measured during sliding between saw cut cylinders of Barre granite deformed in a servo-controlled triaxial loading machine. Two different effects occur, depending on the type of loading. When steady shortening of a stably sliding sample is halted suddenly, slip continues at a diminishing rate, which depends on the logarithm of the initial sliding velocity. Thus, the frictional strength of the surface is decreasing with time. When shortening at a constant rate resumes after the holding period, however, the frictional resistance temporarily rises to a peak proportional to the logarithm of time of no driving, showing a well-known increase in friction with time. The subsequent decrease in resistance from this peak is continuous. The observed frictional resistance of a surface thus results from the interaction of several processes and is not a constant, in general, but depends on the immediate and preceding deformation. When shortening is resumed after a period of holding, slip rate across the friction surface temporarily accelerates above the driving rate. The slip rate can then either decrease to the driving rate and remain stable or accelerate further and lead to stick slip motion, at stress levels below the peak value. There is no apparent difference in the frictional resistance, until the unloading rate of the loading machine is exceeded, between slip episodes which settle stably to the constant driving rate and those which become unstable and result in stick slip. The stick slip failure criterion is not a peak stress but rather loss of strength more rapidly that the unloading rate of the loading machine.The evidence that accelerating motion before stick slip is due to the time variable strenght of rock suggests that accelerating slip will occur before earthquakes. Unfortunately for easy observation, experiments indicate that slip need not be uniform over a fault surface. Also to be established is whether slip will be sufficiently large in the earth to be detectable. Nevertheless, a plausible mechanism for generating observable phenomena prior to earthquakes is suggested by the laboratory observations.