Variance reduction in output from generative AI