Probably an over-representation in the training data really so it's causing overfitting. Because using training data in amounts right from the Internet it's going to be opinionated on human culture (Bart Simpson is popular so there are lots of images of him, Ori is less well known so there are fewer images). Ideally it should be training 1:1 for everything but that would involve _so_ much work pruning the training data to have a roughly equal effect between categories.