
If you train an image recognition system using photos of only light-skinned people, the system will likely not recognize dark-skinned people. This is because of bias in the selection of the inputs.
Likewise if you train a health care AI system using data from only US residents you will likely see a bias from diseases caused by poor diet, e.g. hypertension, heart disease, common cancers (e.g. breast, colon, prostate), and type 2 diabetes. Then if you tried to use that system in African countries the results might be skewed because Africans in general eat less animal products.