Park SocialSimulacra UIST22
The paper introduces the concept of “social simulacra” – large, synthetic populations of users and their textual interactions generated via large language models – as a design tool for prototyping social computing systems. The authors present SimReddit, a system that uses GPT‑3 prompt chains to (1) GENERATE realistic Reddit‑style user profiles and posts, (2) WHATIF designers can edit user attributes or community rules and instantly see updated interactions, and (3) MULTIVERSE allows exploration of multiple plausible community evolutions. Technical evaluation shows SimReddit can recreate the content distribution of 50 real subreddits with high fidelity, while a designer study (N=16) demonstrates that participants can rapidly iterate on community designs, uncover design tensions, and generate richer design concepts compared to manual mock‑ups. The paper also discusses limitations (model bias, lack of multimodal behavior), ethical considerations (potential misuse, offensive content), and situates the work within a broad literature spanning social computing, HCI prototyping, language model prompting, and online community research – as reflected in the extensive reference list. The authors argue that social simulacra can accelerate early‑stage design, enable systematic exploration of design spaces, and complement existing user‑centered methods.
Topics
Social Simulacra: Synthetic Populations for Early‑Stage Design
Introduces the notion of “social simulacra” – large, LLM‑generated synthetic user populations and their textual interactions – as a design tool for prototyping social computing systems. Positions the work within social computing, HCI prototyping, and language‑model prompting literature.