Energy conservation is widely used as a measure of accuracy for molecular simulations. When reporting rates of energy drift, researchers usually assume it is linear in the simulation length, temperature, and system size. We study these assumptions and find that all three are incorrect. Energy drift is too complicated to characterize with a single number, and a more sophisticated analysis is needed to identify the effects of systematic versus random drift, and of integration error versus numerical error. We further argue that energy conservation is not a reliable measure of accuracy. Having small overall drift on long time scales is not a sufficient condition, and in some cases not a necessary condition, for a simulation to produce meaningful results.