Maybe this is more common than I realize, but I was surprised they used Mechanical Turk to find subjects.
It is more common than you realize. Turkers are all over the place in psychology, sociology, machine learning, etc. It's something of a methodological issue: Turkers are extremely convenient and cheap, but they are also not necessarily representative of any realistic population, and the data can be absolute garbage sometimes... but at other times, they're great and they get you results at 1/10th or better the cost, so you can't just write off Turk entirely. Too useful to not use a lot, too dangerous to use blithely.
Are they less representative than other common sources of subjects, like undergrads?
That's difficult to say. Undergrads are at least less likely to be composed of a bunch of bots randomly-clicking on answers and adversarially bypassing your attention checks.