In general, quality is about what you would expect from someone with 2 cents on the line. You pretty much have to write your own logic to pit Turkers against each other for every unit of finished output, or some of the abysmal ones will escape into whatever the project you're doing. (My client just filed 10 bugs against my data set caused by a) turning the floor for data quality down in options and then noticing b) "amusing" Turker understandings of the instructions.)
That said, a few hundred bucks saved me a week of work (and by implication, saved my client several thousand dollars) of having me have to hand-classify a few thousand datums. For their specific need, Turker accuracy was good enough.