PaperlayerView exact revision on arXiv

Source-linked AI summary

Social Simulacra: Creating Populated Prototypes for Social Computing Systems

Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein

arXiv:2208.04024v1cs.HC

TL;DR

Social computing designers lack effective ways to anticipate behaviors that emerge when communities scale beyond small groups. This paper introduces social simulacra, which generate populated communities and varied interactions from design descriptions. Evaluations found that participants often could not distinguish generated from real conversations, while designers used simulacra to refine their designs.

  • Problem

    Small-group social-computing prototypes provide limited evidence about behaviors and challenges that may arise when systems are fully populated.

  • Method

    Social simulacra use large language models to generate users, posts, replies, and anti-social behaviors from a community’s goal, rules, and personas.

  • Results

    Participants misidentified the real conversation in 41% of pairs on average across 50 subreddits, and designers refined designs after using simulacra.

  • Takeaways & Limitations

    Social simulacra let designers examine a breadth of possible social behaviors and explore how design changes or interventions might shift community outcomes.

  • Takeaways & Limitations

    Social simulacra cannot make a single point prediction of how a social system will develop because social dynamics are complex and unpredictable.

Abstract

from arXiv · show

Social computing prototypes probe the social behaviors that may arise in an envisioned system design. This prototyping practice is currently limited to recruiting small groups of people. Unfortunately, many challenges do not arise until a system is populated at a larger scale. Can a designer understand how a social system might behave when populated, and make adjustments to the design before the system falls prey to such challenges? We introduce social simulacra, a prototyping technique that generates a breadth of realistic social interactions that may emerge when a social computing system is populated. Social simulacra take as input the designer's description of a community's design -- goal, rules, and member personas -- and produce as output an instance of that design with simulated behavior, including posts, replies, and anti-social behaviors. We demonstrate that social simulacra shift the behaviors that they generate appropriately in response to design changes, and that they enable exploration of "what if?" scenarios where community members or moderators intervene. To power social simulacra, we contribute techniques for prompting a large language model to generate thousands of distinct community members and their social interactions with each other; these techniques are enabled by the observation that large language models' training data already includes a wide variety of positive and negative behavior on social media platforms. In evaluations, we show that participants are often unable to distinguish social simulacra from actual community behavior and that social computing designers successfully refine their social computing designs when using social simulacra.

1 INTRODUCTION

Social simulacra address the difficulty of anticipating large-scale social behaviors by generating populated communities with diverse interactions. They help designers examine alternative outcomes, test interventions, and refine rules in response to simulated antisocial behavior.

  • Motivation: Small-group experience prototypes may miss behaviors that emerge when a social computing system is fully populated.Designers therefore can be surprised by norms, newcomer enculturation, and antisocial behaviors arising at scale.
  • Introducing Social Simulacra: Social simulacra take community goals, rules, and personas as input and generate many users and textual interactions to populate the social space.The generated interactions include posts and replies intended to reveal both pro-social and anti-social behaviors.
  • Design Iteration: Designers can revise community rules after observing simulated antisocial behavior, such as troll posts, decrease under clearer norms for constructive feedback.The example rules focus critiques on writing rather than the person.
  • Exploring Alternatives: Social simulacra surface multiple possible outcomes rather than a single prediction and support “what if?” exploration of member or moderator interventions.Possible replies include supportive responses and promotional content, illustrating the breadth of behaviors a thread might produce.
  • Evaluation: In a technical evaluation across 50 newly created subreddits, participants misidentified the real conversation in 41% of pairs on average, near chance.The evaluation compared one real and one generated conversation for each community.

2 RELATED WORK

Prior social-computing prototyping struggles to expose the breadth of interactions and edge cases that populated systems can produce. Social simulacra extend prototyping by generating social behavior with language models trained on varied social-media behavior, while retaining a non-predictive scope.

  • Prototyping Foundations: Social simulacra build on interactive prototyping goals of exploring alternatives and proxying user behavior, while presenting generated users and interactions for social-system design.This connects the technique to broader prototyping practices that use approximations to prompt reflection and insight.
  • Social Computing Prototyping: Designers must anticipate interactions among many participants, whose interdependent behaviors create diverse edge cases and potential antisocial outcomes.Examples include trolling, hate speech, and inflammatory comments, which can make systems ineffective or harmful.
  • Social Computing Prototyping: Populating large social systems for prototyping is often an insurmountable challenge because emergent behaviors are difficult to observe before deployment.The core goal is to understand populated-system behavior before people inhabit the space.
  • Social Computing Prototyping: Social simulacra offer a technique for generating content and replies that lets designers reflect on which behaviors their designs are prepared to handle.The approach addresses the scarcity of social-computing prototyping techniques.
  • Large Language Models: Large language models support the approach because they can perform varied language tasks without fine-tuning and are trained on extensive web data containing social-media behavior.This behavioral coverage also means models may generate harmful outputs, including troll behavior, which the system uses as a design input rather than treating solely as an error.

3 SOCIAL SIMULACRA AND SIMREDDIT

Social simulacra use large language models to populate social-computing prototypes with plausible, varied behaviors, helping designers explore how community designs may behave at scale. SimReddit operationalizes this approach through generated communities, intervention scenarios, and alternate outcomes that support iterative design refinement.

  • Scope and trade-off: Social simulacra expand early prototyping beyond small test groups but risk generating behaviors detached from what might occur in practice.The technique addresses the difficulty of recruiting a critical mass while introducing a fidelity trade-off.
  • Generate: SimReddit generates diverse users, posts, and replies from a community’s goal, rules, and target population.Seed personas can be expanded into many non-repeating, thematically relevant personas, while goals shape topics and rules nudge behavior.
  • WhatIf: WhatIf lets designers inspect how individual conversations might change when another persona responds or a moderator intervenes.Designers can select an utterance and request scenarios such as troll responses or specified moderator comments.
  • Design iteration: Designers can use generated interactions to revise community rules and moderation guidance toward focused feedback and away from trollish behavior.Examples show designers changing rules after comparing thoughtful, overwhelming, and hurtful responses.
  • Multiverse: Multiverse produces multiple possible interaction outcomes rather than a single prediction, exposing uncertainty in social-system behavior.Community-wide regeneration resamples persona combinations so designers can compare alternative communities and potential failures.

4 CREATING SIMULACRA USING A LARGE LANGUAGE MODEL

Social simulacra require more than bare prompts: they must encode community structure and diverse, consistent personas to generate useful social-media threads. The approach uses GPT-3 prompt chains to produce varied users and interactions.

  • Prompt chains generate diverse, consistent personas and socially structured posts and replies from community descriptions and seed personas.The prompting technique addresses the complexity that bare prompts do not capture.
  • Social simulacra assume the model can generate the design space’s relevant modality and encode enough world and people knowledge.GPT-3 is presented as a model rich enough to support these assumptions.
  • GPT-3 can produce broad topical outputs from simple prompts, but these outputs are not sufficient by themselves for realistic community threads.Examples include career questions and birthday advertisements.
  • Bare GPT-3 prompts mostly generate generic life updates or advertisements rather than community-structured discussions.Useful threads require top-level posts, engaged replies, and adherence to community rules.

4.2 Prompting Techniques

The prompting pipeline expands a small persona set, generates rule-constrained top-level posts, and iteratively creates varied replies. It also supports WhatIf interventions and Multiverse exploration through persona changes and increased randomness.

  • 4.2.1 Generate – Step 1. Expand on personas.: Designers provide a small set of personas, which the system expands into a larger collective to increase behavioral diversity.The default interface requests ten personas, reducing designer burden while supporting broader generation.
  • 4.2.1 Generate – Step 1. Expand on personas.: The interface supports behavioral rather than demographic personas to avoid stereotyping, although the model can accept gender and race information.The interface does not expose those demographic inputs.
  • 4.2.1 Generate – Step 1. Expand on personas.: GPT-3 expands seed personas with semantically related members, such as additional international-relations students and professors.Generated personas resemble the inputs in meaningful ways.
  • 4.2.2 Generate – Step 2: Generate top-level posts.: Top-level-post prompts combine a persona, community rules, and structured markup so generated content fits the target community.The closure tag signals when comment generation ends.
  • 4.2.3 Generate – Step 3: Generate replies.: Replies are generated iteratively with variable thread lengths, persona turnover, prior-conversation context, and an eight-reply maximum.A per-thread reply probability is sampled around .65, while new personas are selected at a 50% rate.
  • 4.2.3 Generate – Step 3: Generate replies.: Reply prompts reiterate the responder’s persona and community rules, producing responses aligned with those constraints.The example generates a supportive reply to a post about antidepressants.
  • 4.2.4 WhatIf.: WhatIf replaces the current responder’s persona to simulate requested behaviors, such as trolling and responses to moderator intervention.The troll example generates an explicitly anti-antidepressant reply.
  • 4.2.4 WhatIf.: Multiverse increases GPT-3’s temperature from 0.7 to 0.8 while keeping the prompts unchanged to obtain more varied outputs.Temperature controls generation randomness, from deterministic at 0 to highly random at 1.

4.3 GPT-3 Settings

The implementation uses GPT-3’s base davinci engine and exposes model-generation settings relevant to reproducibility and knowledge scope.

  • The studies used GPT-3’s base davinci engine, released June 11, 2020, rather than later davinci variants.Consequently, the model lacks knowledge of world events occurring around or after that date, including the COVID-19 pandemic.

5 TECHNICAL EVALUATION

The technical evaluation tested whether SimReddit could reproduce plausible behavior in 50 post-training subreddits and whether prompting components mattered. Participants often struggled to distinguish SimReddit conversations from real ones, while ablations were less plausible and some generations lacked needed domain knowledge.

  • The evaluation criteria required generated topics and behaviors to resemble what might arise after deployment rather than remain generic or irrelevant.The study therefore assessed realism and design usefulness through staged evaluations.
  • The technical evaluation repopulated 50 English subreddits created on or after GPT-3’s release, limiting direct repetition of training data.The sample was selected using subreddit creation dates and subscriber counts.
  • Participants compared one real conversation with one synthetic conversation per subreddit after browsing the target community’s topics and behaviors.Conditions included crowdworkers, SimReddit, and SimReddit ablations without community descriptions or personas.
  • 41% (std=10%) was the average error rate for distinguishing real conversations from SimReddit conversations, versus 50% for random guessing.Participants correctly identified generated content only slightly more often than they were wrong.
  • SimReddit significantly outperformed all ablations, including 21% (std=15%) without community descriptions and 33% (std=10%) without personas.The omnibus test reported F(3, 196) = 22.49, p < 0.001, with pairwise comparisons at p < 0.01.
  • At its best, SimReddit reconstituted domain knowledge in social contexts and surfaced problematic behaviors such as antivaccination persuasion.The examples include Cyberpunk 2077 discussion and vaccine-related interactions.
  • Generations were not always plausible: some lacked domain knowledge and others became unexpected, generic, or off-topic.The COVID-long-haulers example lacked enough information for a meaningful generation, while an ablation produced an unrelated political question.

6 DESIGNER EVALUATION

This section evaluates whether SimReddit helps social computing designers iterate and improve their designs by turning its insights into concrete inspirations for a new subreddit.

  • The evaluation asks whether SimReddit helps designers iterate and improve their social computing designs.
  • It examines how insights provided by SimReddit materialize into concrete design inspirations.
  • Participants are tasked with designing a new subreddit as the setting for the evaluation.

6.1 Study Procedure

Participants designed and iterated on a subreddit using SimReddit generations, revised their designs after reviewing simulated conversations, and explored troll interventions through WhatIf and Multiverse.

  • Participants designed a new subreddit community, drafting and iterating on its description, rules, and comments against a troll.
  • The study used a screener, a pre-interview design task, and a video-call interview.
  • Researchers discussed participants’ prior design challenges, subreddit goals, and intended community topics and behaviors.
  • After reading generated conversations for 5 to 10 minutes using think aloud, participants revised their designs and could trigger a new SimReddit generation.
  • WhatIf and Multiverse presented three troll-response generations, prompting participants to consider surprises and moderator interventions.

6.2 Participants

The study recruited 16 people with prior experience designing or moderating online social spaces and involved sessions lasting 90 to 120 minutes across multiple days.

  • The study recruited 16 participants with prior experience designing or moderating online social spaces.
  • Participants were recruited through social media and mailing lists.
  • The study lasted 90 to 120 minutes spread across multiple days, and participants were paid $50.00.

6.3 Inductive Analysis of the Interview and Participants’ Designs

Researchers analyzed interviews and participant designs inductively, moving from prompt-level codes to higher-level themes about unexpected behaviors and design improvements.

  • The researchers followed the same inductive analysis procedure as in their technical evaluation study.
  • Initial codes closely reflected prompt-sentence characteristics, including desirable generated behaviors and added design rules.
  • The analysis synthesized codes into themes such as unexpected content or behavior and improvements resulting from design changes.

6.4 Results

SimReddit gave designers unexpected positive, negative, and borderline examples that sharpened their understanding of possible community behavior. These generations prompted nearly all participants to revise their designs, while participants generally found them realistic but recognized limits to prediction.

  • Generations offer concrete design insights: Generated communities revealed desirable behaviors beyond participants’ expectations, such as members seeking friends to attend Pittsburgh events together.This helped one participant recognize additional value that their community could provide to students living around Pittsburgh.
  • Generations offer concrete design insights: Negative and borderline generations prompted reflection on moderation and rules, including vigilance against trolling and debate over whether vague posts should be allowed.Participants used these cases to identify risks and weigh competing goals for productive discussion.
  • Iteration improves the community: 15 of 16 participants revised their original design, aiming either to prevent failure cases or to encourage particular community cultures and norms.Examples included rules against promotional content or conflict, and aspirations for happy or non-creepy communities.
  • Iteration improves the community: 10 participants were largely pleased with the content changes produced by their revisions and connected them to a better understanding of what their rules should include or exclude.Participants described updated parameters as shifting the generated community closer to their vision.
  • Role in the design process: 14 participants described the generations as generally realistic, but 15 noticed implausible aspects and questioned whether rule changes could predict likely human behavior.All 16 participants nevertheless said SimReddit added value to how they thought about and crafted social designs, especially by grounding assumptions and supporting moderator discussions.

7 DISCUSSION

Social simulacra broaden social-system prototyping by exposing plausible behaviors and supporting proactive design, while requiring expert interpretation because they cannot exhaustively predict outcomes. The approach also carries technical, ethical, and societal risks that constrain deployment and generalization.

  • Design implications: Social simulacra help designers question assumptions and explore a broader design space, but must be coupled with expertise to judge which behaviors warrant action.They may cue recall more effectively than checklists and support nuanced discussions of problematic behaviors.
  • Design implications: Prototyping with social simulacra can shift social-computing design toward proactive evaluation before deployment, reducing reliance on learning only after users experience harm.Interviewees described reactive strategies as requiring harm to users and communities before improvement.
  • Limitations: Social simulacra are unlikely to enumerate every possible social outcome, creating blind spots and a risk that generated issues falsely increase confidence that no others exist.Their breadth may nevertheless exceed what small groups of test users can provide.
  • Limitations: Because social dynamics are complex and unpredictable, social simulacra cannot make a single point prediction of how a social system will develop.The paper frames presenting many outcomes as an open design space rather than a forecasting capability.
  • Limitations: SimReddit remains limited by prompt length, text-only output, and English-language operation, while future multimodal models could broaden supported social spaces and content.GPT-3’s roughly 8,000-character prompt limit can require truncating conversation history and prevents conveying a broader set of posts.
  • Ethical and societal impact: Generated content can reproduce biased, hateful, or upsetting behavior, while the technique also risks enabling astroturfing, harassment attacks, and propaganda.The paper recommends vetted users and centralized hosting with logging so generated content can be audited.

8 CONCLUSION

The conclusion presents social simulacra as a way for social-computing designers to envision behaviors in populated spaces. The generated reviewers offer differing assessments, praising the technique’s potential while identifying insufficient detail and limited discussion of benefits and limitations.

  • 8 CONCLUSION: Social simulacra enable social-computing designers to envision behaviors that might occur when their space is populated.The conclusion illustrates the approach with a SimReddit community dedicated to reviewing UIST papers.
  • 8 CONCLUSION: Other generated reviewers criticized the paper for insufficient concrete detail and requested more discussion of the approach’s benefits and limitations.These comments identify presentation and coverage concerns rather than disputing the technique’s stated purpose.
  • 8 CONCLUSION: One generated reviewer praised the technique as useful to designers and called it a significant contribution to social computing.The reviewer also described the paper as well-written and clearly illustrated with examples.
Loading 2208.04024v1…