When people say they're going to vote, do they really vote? Not according to this study (warning, not yet peer reviewed). Even worse, from a research standpoint, they find many who say they're not going to vote actually voted.
The study gets at the question a number of ways. One is simple. Look at who said they're gonna vote in a survey and then go to the courthouse and see if they actually voted (it's a public record whether or not you voted, but not who you voted for). They also used an experiment and some other approaches.
To summarize, these "surprisingly inaccurate self-predictions" are full of measurement error, which can lead to all kinds of problems when you rely on them to, oh, decide on "likely voters" in a horserace poll (this is "grossly inadequate," the authors say) or even as a measure of political interest, motivation, or intent in academic studies. The authors argue self-prediction is "far inferior to past vote history as a predictor of actual turnout."
Again, the caveat -- this manuscript is under review somewhere and has not as yet undergone a complete peer review. As such, take the results with a grain of salt until it's been vetted by folks who are serious about this stuff.
But, there's some fascinating discussion at the end of the paper about the psychological underpinnings of saying you're gonna vote and you don't, or to me even more fascinating, saying you're not gonna vote and you do it anyway. That latter, they argue, may in part be a function of frustration with the political process but among regular voters, the pull of Election Day is simply to much. So pollsters in particular may be losing important data with their traditional screening questions.