Because said set of numbers is produced via a training process that has the original as an input, and a different input would produce a different set of numbers.

You're correct that merely containing the information would not violate copyright - it's all about how that information was produced.