An imaging algorithm that scores well on its own test set may still fail in the clinic. It may have learned a shortcut in the data and not the disease. It may work for some groups of patients and not for others. It may be upset by a new scanner or a change in the population, and its probabilities may mean something different where disease is rarer. This chapter deals with the four properties that decide whether an algorithm can be trusted: whether its decisions can be explained, whether it is fair, whether it is robust, and whether it has been properly validated. It describes what explanation methods show and what they do not, the sources of bias along the path from patient to prediction, the effects of dataset shift and deliberate perturbation, calibration, the ladder of evidence from internal testing to clinical trials, and the reporting standards that now apply. Simulations computed for the chapter illustrate a shortcut, an under-represented group, a deliberate perturbation, miscalibration and the gap between internal and external performance. The final sections cover monitoring after deployment and the questions to ask of any product.