A South Korean man has been sentenced to jail for using artificial intelligence to generate exploitative images of children, the first case of its kind in the country as courts around the world encounter the use of new technologies in creating abusive sexual content.
That is not required. Especially in the larger models like a DALLE-3 it can combine concepts even without being directly trained on it. The one they had in the showcase for DALLE-2 was a chair shaped like an avocado. It knows what a chair is and it knows what an avocado is, so it can combine them. So it can know “this is what a naked human looks like” and “this is what a human child looks like” and could combine them without having ever seen CSAM.
Exxept for the source csam required to get the model started, of course.
That is not required. Especially in the larger models like a DALLE-3 it can combine concepts even without being directly trained on it. The one they had in the showcase for DALLE-2 was a chair shaped like an avocado. It knows what a chair is and it knows what an avocado is, so it can combine them. So it can know “this is what a naked human looks like” and “this is what a human child looks like” and could combine them without having ever seen CSAM.
Did somebody audit the dataset?
Ya, I thought so…