Apple's description of the training process (https://www.apple.com/child-safety/pdf/CSAM_Detection_Technical_Summary.pdf https://www.apple.com/child-safety/pdf/

Apple's description of the training process (https://www.apple.com/child-safety/pdf/CSAM_Detection_Technical_Summary.pdf https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...) sounds like they're just training it to recognize some representative perturbations, not useful semantic categories.