Abstract
Evaluating personalized artificial intelligence (AI) systems may be challenging, as the use of detailed user context can conflict with user privacy considerations. A system is described for generating synthetic user populations via a world simulation. The system can create generative agents with demographic and behavioral characteristics based on aggregated, anonymized statistical data. These agents can interact over time within a simulated environment, producing a stream of longitudinal user behavior captured in isolated test accounts. The resulting dataset can be used for scalable, automated testing and debugging of AI models, offering a contextually rich, privacy-respecting alternative to using real user data. This approach allows for detailed inspection of synthetic user histories to help diagnose model failures.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Venkatanarayan, Amey; Liu, Yen-Ting; Kodur, Kamakshi; Flynn, Paddy; Hu, Yue; Jimenez-Olivas, Jose; and Song, Zhewen, "Generating Synthetic User Populations via World Simulation for Privacy-Safe AI Evaluation", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10819