This paper describes the effects of input saturation in model-based fault diagnosis. A linear model has been used for the residual generator, in order to show its lack of robustness. The experiments have been developed on a simple simulation case: a liquid tank. The aim of the paper is to show the relation between the degree of nonlinearity and the performance of a diagnosis method based on linear model. For this purpose, only additive faults have been considered. It has been taken into account sensor and actuator noises and we have considered the system works far from operating point too. It is shown that saturation presence decreases detectability of additive faults.