So is the car manufacturer responsible if someone drives their car into the sidewalk to kill some people?
Your analogy doesn't match the premise. (Again assuming there is no csam in the training data which is unlikely) the training data is not the problem it is how the data is used. Using those same picture to generate photos of medieval kids eating ice cream with their family is fine. Using it to make CSAM is not.
It would be more like the doctor using the nazi experiments to do some other fucked up experiments.