I think you’re suggesting that, if it was trained on adult sexual abuse material and images of children, it could composite them in a way that made images depicting CSAM. Doing this would be far different to putting Luke Skywalker on a kangaroo as children an adults look fundamentally different, and without source material it couldn’t generate that.
people keep saying this, and, no
you can go on nano banana and ask for a picture of kim novak punching bigfoot with luke skywalker riding a kangaroo in the background
do you think nano banana was trained on photos of kim novak punching bigfoot with luke skywalker riding a kangaroo in the background
no
it knows those things individually and can composite them together sorta in a way that imitates the layout of other miscellaneous photos
I think you’re suggesting that, if it was trained on adult sexual abuse material and images of children, it could composite them in a way that made images depicting CSAM. Doing this would be far different to putting Luke Skywalker on a kangaroo as children an adults look fundamentally different, and without source material it couldn’t generate that.