I contend that is impossible to make an unbiased AI. I did an AI image recognition project several years ago. It used yolo to categorize rust into grade 1, 2, and 3 for offshore platforms. When creating our training dataset, we had different rust inspectors from different parts of the world drawing different lines in the sand between what was category 1, 2, and 3. We had to eventually pick which bias we wanted to roll out worldwide. The advantage for a giant corporation was that now the same consistent bias was being used worldwide and fewer people had to be safety trained to go on the offshore platforms. If that incredibly dull and basic application can’t be unbiased, I don’t think it is possible to avoid bias in anything produced with a training dataset. The very word “training” implies it. Someone somewhere decides A is in the training and B is not, and a bias is born, intentionally or not.

So the task is really to find the AI with the bias that works best for your application, not to try and remove bias.