AI with no dibiasing does not work often enough. I’ve seen enough examples in Computer Vision models to say so. (Even segmentation models tend to rely on external clues and this affects the generalizability of the model). Happy to share more via email.