>> how would you create a compelling CSAM image with nothing resembling the target in the training data?
(1) You feed it extra training data to suit the desired output.
(2) The AI combines bits from existing training data. It has non-pornographic content of children. It has adult pornography. Layer one over the other and voila. Crime. Most AI has zero images of cats doing calculus, but is very capable of generating such content.
(3) You manually aid the process. People forget how powerful standalone AI can be when one manually selects good output and feeds it back as a guide for the next iteration. This does not scale to a multi-task environment, but is very useful when accomplishing one specific task.
In theory, without the adult pornography to help seed step 2, how hard would you think it to still accomplish this? ie: what if we somehow target not-training LLMs on any pornographic material or any material with children?
(1) You feed it extra training data to suit the desired output.
(2) The AI combines bits from existing training data. It has non-pornographic content of children. It has adult pornography. Layer one over the other and voila. Crime. Most AI has zero images of cats doing calculus, but is very capable of generating such content.
(3) You manually aid the process. People forget how powerful standalone AI can be when one manually selects good output and feeds it back as a guide for the next iteration. This does not scale to a multi-task environment, but is very useful when accomplishing one specific task.