Fairness Regularization Fails to Improve Demographic Parity in Legal BERT

A study on LegalBERT shows that penalizing stereotypical warmth and competence representations during fine-tuning fails to reduce demographic disparity on the LexGLUE ECtHR corpus. The results indicate that debiasing interventions do not necessarily improve demographic fairness despite changes to model representations.