The paper introduces social simulacra, a prototyping technique that uses large language models (LLMs) to generate realistic, large‑scale mock communities—including user personas, posts, replies, and anti‑social behaviors—based on a designer’s specification of a community’s goal, rules, and target population. Implemented as the SimReddit tool for Reddit‑style subcommunities, the approach features three core interactions: GENERATE (automatically creates thousands of diverse personas and populates the community with threaded discussions), WHATIF (allows designers to explore alternative responses or moderator interventions for specific utterances), and MULTIVERSE (produces multiple alternative community outcomes by resampling model randomness). The authors detail prompting chains that expand seed personas, generate top‑level posts, and iteratively produce replies while respecting community constraints, leveraging GPT‑3’s knowledge of social media behavior. Two evaluations assess realism and design utility: a technical study shows participants struggle to distinguish simulated from real Reddit conversations (error rate ~41%, near chance), and a designer study with 16 social‑computing practitioners demonstrates that SimReddit helps uncover unexpected positive and negative interaction patterns, prompting rule revisions and moderation strategies. The discussion reflects on how social simulacra extend experience prototyping to social systems, enable proactive design, and expose designers to a broader range of potential harms. Limitations include incomplete coverage of possible outcomes, reliance on English‑only GPT‑3, and lack of multimodal content. Ethical considerations address the generation of harmful content (used as a design aid) and the risk of misuse for astroturfing, recommending vetted access and audit mechanisms. The work concludes that social simulacra provide a scalable, low‑cost method for designers to anticipate socio‑technical dynamics before deployment. The paper situates this contribution within a broad literature on online community behavior, harassment, LLM capabilities and risks, and prototyping methods, citing works ranging from critical mass theory of interactive media to recent studies on fake‑news perception and language‑model ethical risks.
Topics
Social Simulacra Concept and Foundations
Introduces the notion of social simulacra—a prototyping technique that leverages large language models to generate realistic, large‑scale mock communities (personas, posts, replies, and anti‑social behaviors) from a designer’s specification of community goals, rules, and target population. Positions the work within prior research on online community dynamics, harassment, critical mass theory, and LLM risk literature.