The 1985 prediction of a characteristic magnitude 6 Parkfield earthquake was unsuccessful, since no significant event occurred in the 95% time window (1985–1993) anywhere near Parkfield. The magnitude 6 earthquake near Parkfield in 2004 failed to satisfy the prediction not just because it was late; it also differed in character from the 1985 prediction and was expectable according to a simple null hypothesis. Furthermore, the prediction was too vague in several important respects to meet the accepted definition of an earthquake prediction. An event occurring by chance and meeting the general description of the predicted one was reasonably probable. The original characteristic earthquake model has failed in comprehensive tests, yet it is still widely used. Modified versions employed in recent official seismic hazard calculations allow for interactions between segments and uncertainties in the parameters. With more adjustable parameters, the modified versions are harder to falsify. The characteristic model as applied at Parkfield and elsewhere rests largely on selected data that may be biased because they were taken out of context. We discuss implications of the 2004 event for earthquake prediction, the characteristic earthquake hypothesis, and earthquake occurrence in general.