If one only trains the model to correctly identify a disease (e.g. maximising the correct identification and minimising the error), then there is no space for t

If one only trains the model to correctly identify a disease (e.g. maximising the correct identification and minimising the error), then there is no space for the model to be "racist". I don't question that there is a signal inside the images, but someone has to change the model and training to then maliciously use that information.