What if I look at other people's art and learn from it? Seem unfair to pass that work off as my own. All artists should be banned from looking at copyrighted images, we can't risk them incorporating copyrighted elements into their own work. /s
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.
It's still a computer program that uses an enormous amount of copyrighted work as its input.