The SPS specification[1] says you should expect 95% of reported locations to be within a 9m diameter circle (ref 3.8.3). Realistically, though, you can get much better, and a lot of receivers get within 2-3m.
Up until about 2000, there was thil thing called Selective Availability[2] (aka SA). This limited civilian performance to much worse (95% ~20-30m) by some clock dithering techniques. During-SA and post SA accuracies are very different.
In order to support airplanes using GPS for precision landing, the FAA rolled out WAAS[3], which is a wide-area augmentation system that broadcasting atmospheric correction data (specialized just for the US; there are other national systems). GPS+WAAS gets you to within a meter or so. [3]
If absolute position isn't important, but relative position is (say, if you're measuring the movement of two tectonic plates relative to each other at a fault), there's this neat thing called RTK. This gets you relative positions (in the field) on the order of millimeters[4].
[1] http://www.gps.gov/technical/ps/2008-SPS-performance-standar...
[2] http://en.wikipedia.org/wiki/Error_analysis_for_the_Global_P...
[3] http://www.navipedia.net/index.php/WAAS_Performances
[4] http://hemispheregnss.com/DesktopModules/EasyDNNNews/Documen...
There used to be an additional limitation called "selective availability" where the military deliberately introduced errors in the clock siginal to reduce accuracy to about 100m, but this was switched off in 2000.
http://www.gps.gov/systems/gps/performance/accuracy/
I can personally say that ~1.5-2m accuracy is definitely achievable even with a cheap entry-level receiver like the ones integrated into smartphones, when there is a clear view of the sky and good weather.
Edit: It seems my phone does have DGPS, so that accuracy is not just "raw GPS".