It's a fine analogy, but the map is not the territory. Machine learning is not human learning, even if it works in a vaguely comparable way.

It's still a computer program that uses an enormous amount of copyrighted work as its input.

"With four parameters I can fit an elephant, and with five I can make him wiggle his trunk”

It seems like you could calculate how much data is within X% error of a 5GB model, and what X% should be for 'visual data'.

I bet it's pretty big.