Die University of Utah and Prof. Dr. C. Obvious wrote a paper (PDF):
Once police are deployed based on these predictions, arrest data from the neighborhood is then used to further update the model. Since arrests only occur in neighborhoods that police have been sent to by the predictive policing algorithm itself, there is the potential for this sampling bias to be compounded, causing a runaway feedback loop. Indeed, Lum and Isaac have recently shown that this can happen.
To be fair, the paper is about how to remove this kind of bias from precrime systems. On the other hand, predictive systems using past data to predict future actions are obvious sources for feedback messing with the prediction. One would expect this to be taken into account from the start.