Towards Effective Discrimination Testing for Generative AI