An effective approach to homogenize traffic flow is to improve the distribution of traffic demand across the network, known as dynamic routing or load balancing. Load-balancing strategies can be applied by traffic management centers or result from individual route choice decisions. As an instructive example, we apply macroscopic and microscopic traffic flow models to study the system dynamics of a simple networkNetwork of two alternatives during a rush-hour where a fraction of drivers make an active route-choice decision. Naively applying a greedy deterministic routing taking into account the current situation (instantaneous travel times) generally worsens the system performance as soon as a sufficient proportion of drivers is involved. Unlike weather forecasting, dynamic routing and short-term traffic predictions face a conceptual challenge: the prediction undermines itself. Nowadays, such situations are realistic as more and more drivers use real-time traffic information and navigation applications. We also show that introducing a simple form of uncorrelated stochasticity can avoid these negative effects. A time uncertainty of about one minute is sufficient to bring the system near the system optimum although the decision is yet greedy and based on instantaneous criteria.