Free access to a generative AI (GenAI) service requires costly compute, yet can also produce data that improve future service quality. We study a service provider whose users differ in request value and privacy preference, under the constraint that paid requests are kept private, while free requests can be used to improve model quality which otherwise reverts toward a baseline. In particular, we characterize service demand and prove that quality converges to a unique steady state for every stationary service strategy. Next, we analyze optimal pricing and quality policies and show that these can be expressed using two endogenous user value thresholds. For uniform values, we provide a closed-form optimal service strategy and characterize the sufficient conditions for offering free services as dependent on inference cost.

翻译：暂无翻译