> A ML model is clearly a derivative work of its input.

Do you mean this in a copying sense or a mathematical sense?

What if it's only storing 1 byte per input document?