偏见

Systematic skew or unfair tendencies in model output.

Bias mainly comes from imbalanced training data and annotator subjectivity, showing up across gender, region and race. Mitigations: data balancing, debiasing fine-tuning and output review.