Codecs like AAC apply psychoacoustic rules that were only correct for the overall piece being encoded, not samples taken from it. If a song has two overlapping sounds and the human ear physically can't hear both, a lossy codec looks for that and conserves space by completely omitting the sound that was masked, leaving you with no way to recover it if your postprocessing takes away the sound that had been masking it. A lossless codec preserves the sounds you might be able to hear after sampling or remixing, though of course it takes more space to do that.
