Detecting and Mitigating Bias by Treating Fairness as a Symmetry Operation

In three linesPaper formalizing bias as symmetry breaking: a classifier is fair if outputs remain invariant when switching a sensitive attribute while holding merit features fixed. Loss-based regularization restores symmetry. Results: 90% violation reduction, ~5% accuracy cost on 4 synthetic datasets.

Summary generated by Claude — human-verified