I think Thomas meant that designing a lossy compression scheme inherently involves making decisions about what parts of the data it is OK to lose -- i.e., in audio processing it might be acceptable to drop frequencies outside of the range of human hearing from the compressed output.
