Bias refers to systematic prejudice in AI outputs, often caused by unbalanced or non-diverse training data.