Prejudice built into an AI model. An AI model is only as accurate as the data used to train it. As a result, a model can be neutral, as it should be, or highly prejudiced. If enough stories about the world being flat are fed into an AI model, it will be reluctant to ever claim the world is round.
The labeling of training data by humans is also important because they can inject their own prejudice without even realizing it. In addition, the algorithms used to develop the models, as well as those in the inference execution, can exaggerate comments and opinions in the training data. See
AI training vs. inference and
AI programming.