What Fairness Changes in Synthetic Data? A Controlled Audit

We hold a tabular GAN backbone fixed and vary only the fairness term, isolating what a parity constraint does from what an architecture does. Fairness parity on the synthetic table does not reliably transfer to the downstream model. The cost lands on subgroup representation and on one group’s utility. We also introduce α, computable before any model is trained, to show which group’s outcome distribution changes more when parity is imposed.