I should state for the record that I'm not a big fan of polls. I don't think they're very trustworthy, and they appear to be less so with every passing day.
As such, I thought I would pass along this story by David Price at Counter Punch:
As a social scientist who uses both quantitative and qualitative research methods, I am not one to dismiss survey research out of hand, but something has methodologically gone awry when polls are swinging about this wildly. Obviously past telephone-survey presidential polls have accurately predicted election outcomes, but Americans' social interactions via telephones may be evolving in ways that render past telephonic sampling techniques unreliable.
We Americans simply don't answer our phones like we used to. Entire industries are now devoted to helping us not answer the phone. Voicemail, Caller ID, caller-specific-rings, cell-phones, even email have fundamentally transformed the ways we (don't) answer the phone when it rings. These and other technological innovations have moved us from a late-20th Century near-pavlovian automatic response of answering the phone when it rang, to new levels of screening or ignoring calls without a sense that we might be missing something important. When pollsters call under these technological conditions they are now increasingly treated as any telemarketer or unknown caller would be, thus the people who pollsters actually get to talk to are becoming increasingly less representative of the general public. There now may be something unusual about people who are willing to answer the phone a talk with strangers, and we should be skeptical about generalizing from the results of these surveys. It is possible that the new habit of non-phone-answering is evenly distributed throughout the population (thus reducing this as a sampling confound), but this seems unlikely.
I also think phone/internet/whatever polling misses out on people who are more likely to vote Democratic (ie, working poor). Now that may be alright for marketing research where they only care if the respondent can afford whatever they are selling, but the same criteria can not be used when it comes to voters - who likely can not afford another 4 years of whatever it is GW Bush is selling.