AI 中文总结
该研究针对用户请求价值与隐私偏好存在差异的生成式AI服务，刻画了服务需求与模型质量稳态，提出基于用户价值阈值的最优定价和质量策略，明确提供免费服务的充分条件。
AI 中文摘要
生成式人工智能（GenAI）服务的免费访问需要高昂的计算成本，但也能产生可提升未来服务质量的数据。我们研究了一家服务提供商，其用户在请求价值和隐私偏好上存在差异，约束条件为付费请求会被保密，而免费请求可用于提升模型质量，否则模型会回归基线状态。我们刻画了服务需求，并证明对于每一种稳态服务策略，模型质量都会收敛到唯一的稳态。接下来，我们分析了最优定价和质量策略，表明这些策略可通过两个内生用户价值阈值来表达。对于均匀价值，我们提供了闭式最优服务策略，并刻画了提供免费服务的充分条件，该条件取决于推理成本。
英文摘要
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.
Comments39 pages, 1 figure
