In the Netherlands the current dike design policy is to design flood defence structures corresponding to an agreed flooding probability with an extra safety board of at least 0.5 m. For river dikes a return period of 1,250 years is used to determine the design water levels. A problem with this strategy is that it builds on assumptions with regard to the intrinsically uncertain probability distributions for the peak discharges. The uncertainty is considerable and due to (1) the measuring records that are limited to about 100 years and (2) the changing natural variability as a result of climate change. Although the probability distributions are regularly updated based on new discharge data the nature of the statistics is such that a change in the natural variability of the peak discharge affects the probability distribution only long after the actual change has happened. Here we compare the performance of the probabilistic dike design strategy with the older strategy, referred to as the ‘self-learning dike’. The basic principle of the latter strategy is that the dike height is kept at a level equal to the highest recorded water level plus a certain safety margin. The two flood prevention strategies are compared on the basis of the flooding safety over a 100-year period. The Rhine gauge station at Lobith serves as case study. The results indicate that the self-learning dike performs better than the probabilistic design in terms of safety and costs, both under current and climate change conditions.