They act like any other P2P user and download some number of files (i.e. their threshold for selecting victim) and play them. Presumably mp3 content, other then audio, will give them a clue as to what the audio is supposed to be.

They could also use some spectral analysis on the audio to develop a signature to match against a reference set of such signatures of known copyrighted pieces. This would automate the testing and might be followed by validation by a life listener before legal action.

As I said, just a guess on my part.