I'm torn between the statistical arguments for vote-count manipulation and an explanation of the exit poll discrepancies I'll call "the evil voter hypothesis." Perhaps instead I should call it "the preference for an obviously evil candidate hypothesis." It goes like this:
Most voters were not blind to the lies and worse practiced by Bush and his crew. They knew he was a somewhat-evil guy - and that Kerry, morally and ethically, was the better man. But, just as in dating, often the known-evil character can for some reason or another be more attractive. Even so, it's considered rebellious to say, "I like a bit of evil in a (wo)man." It's more normal to present the social pretense of preferring the obviously more virtuous (wo)man rather than the cheat and the snake.
What would make this election unusual, on this hypothesis, is that there was actually a degree of concensus on one candidate being more evil than the other, rather than the more normal situation (say Carter-Ford, or Reagan-Carter or even Bush I-Clinton) where the concensus is that both men are of better-than-average character by standard social norms.
But, in the wake of 9/11, there's a willingness on the part of much of the public - not most perhaps but a significant minority - to do evil in the world, as expressed in the toleration of torture and unlawful detention and so on. But few men in church on Sunday will admit that their favorite part of the week is Friday night out drinking with the whores and lynching the niggers - even though when it comes to their monetary donations, affection and enthusiasm, their spirits are more with the whores and lynch mob than some namby-pamby Jesus.
So the public exit pollsters were lied to by these people just for the same reason that, well, Clinton denied "sex with that woman"; most of us know and generally follow the rule of not admitting a preference for the obviously evil in public. And the election was unique in having evil predominantly on one side of the choice.