Think of all the use cases where the output is going to be ingested by another machine. You don't know that "perceptually lossless" as designed for normal human eyeballs on normal screens in normal lighting environments is going to contain all the information an ML system will use. You want to preserve data as long as possible, until you make an active choice to throw it away. Even the system designer may not know whether it's appropriate to throw that information away, for example if they're designing digital archival systems and having to consider future users who aren't available to provide requirements.