e-hadj's writeup is very, very good but he makes a mathematical mistake and ignores the ever-present tracking polls.
First of all, the minimum swing detectable is actually closer to 4.5%, because you don't just add up standard deviations, you add up the squares and take the square roots. Also, even with a 4.5% swing, you're only 95% confident that there's an actual swing--i.e. there's a five percent chance that there was no swing.

Now, onto the more interesting feature of this presentation: tracking polls.
In recent Presidential elections, companies like Gallup and Zogby set up tracking polls. Unlike regular polls, which are released once a week or even less often than that, tracking poll results are released every day. Unfortunately, due to cost and other issues, polling companies do not want to have to call more than 300 or so people per day, well below the 1000 or so required for the desired 3% margin of error. So what do they do? The three-day rolling average.

The three-day rolling average is simple: Instead of just publishing last night's results(opinion polls are conducted in the evening, when people are most likely to be home, and released in the morning, snap polls being a major exception), polling companies take last night's results, the results of the night before, and the results of the night before that and take the mean of the values. A side-effect of this is that if you subtract consecutive poll results from a tracking poll, you are actually subtracting one night's result from the result three nights prior to that. This means that if you want to get an accurate picture of how big an opinion shift is you should wait at least three days for the new results to sink in. It also means that a particularly good night for one candidate can linger for three days and create "phantom shifts" in the rolling average, creating an unreasonably large bump in the opposite direction once the "good night" drops off the rolling average.

All of this is why you should never look at just one poll, but instead at a variety of polls. Sites like fivethirtyeight.com are (or rather, were, and will be next election cycle) very useful for this purpose, since it averages together a collection of polls from state and national sources.

In conclusion: If somebody is on cable news talking about polls, and his name isn't Nate Silver, he's probably overstating his case.