Fairness as symmetry — new framework cuts AI bias 90%
arXiv paper formalizes bias as symmetry breaking: a classifier is fair if flipping a sensitive attribute (race, gender, age) leaves predictions unchanged when merit features stay fixed.
Loss-based regularization restores symmetry. On four synthetic datasets, the framework achieves 90% bias violation reduction with ~5% accuracy trade-off.
No causal graph required, lightweight to compute, works for any binary sensitive attribute. Practical for high-stakes deployment (hiring, lending, criminal justice).