I'm guessing the rise of generative AI will help ensure this never happens again.
If you wanted to build a Sinatra Simulator, you'd want as many takes as possible to train it. For that reason, I imagine any and all high resolution recordings are now considered gold rather than something you just keep around for old times' sake, possible use in reissues, research, etc.
Even session background singers or isolated instrument tracks would be valuable; with all kinds of training data / not-samples, producers could add bits of Quentin Tarantino-esque pastiche to tracks easily -- "give me some background singers that sound like the ones in Smokey Robinson's first album..." "I want this bass track to sound like Duck Dunn playing a Hofner Beatle Bass..." Training data is what you need to make that happen, and record company vaults are the mines where that gold comes from.
If you wanted to build a Sinatra Simulator, you'd want as many takes as possible to train it. For that reason, I imagine any and all high resolution recordings are now considered gold rather than something you just keep around for old times' sake, possible use in reissues, research, etc.
Even session background singers or isolated instrument tracks would be valuable; with all kinds of training data / not-samples, producers could add bits of Quentin Tarantino-esque pastiche to tracks easily -- "give me some background singers that sound like the ones in Smokey Robinson's first album..." "I want this bass track to sound like Duck Dunn playing a Hofner Beatle Bass..." Training data is what you need to make that happen, and record company vaults are the mines where that gold comes from.