CoChat
Sign Up
Cited by 212Innovative Human-Technology InteractionUsability and User Interface DesignRecommender Systems and Techniques

Social Simulacra: Creating Populated Prototypes for Social Computing Systems

Open access
Park, Joon Sung, Popowski, Lindsay, Cai, Carrie, Morris, Meredith Ringel, Liang, Percy, Bernstein, Michael S.·2022·Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology
DOI 10.1145/3526113.3545616
Open this paper in CoChat — free accountOpen PDF

Abstract

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.

Integrity signals

Registry

Crossref · 10.1145/3526113.3545616

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 36 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Using social simulacra, participants found positive use cases they had not previously considered.

    “With social simulacra, participants identifed positive use-cases they had not considered (e.g., impromptu friend-seeking to go sightseeing in a community for sharing fun events around Pittsburgh) and negative behaviors that they had not accounted for”
    Quote in the paperSubstantial claim
  • Findingreasoning

    Using social simulacra, participants discovered negative behaviors they had not anticipated.

    “and negative behaviors that they had not accounted for (e.g., Russian trolls shifting the tone of an international afairs discussion community).”
    Quote in the paperSubstantial claim
  • Findingnone given

    Even experienced designers felt overwhelmed when trying to imagine the interactions that might occur in their design.

    “Even seasoned designers found it overwhelming to envision the possible interactions that could take place in their design, and as a consequence, were in the practice of waiting until problems emerged and their communities were damaged to add rules and interventions.”
    Quote in the paperSubstantial claim
  • Findingown data

    In the technical evaluation, participants frequently could not tell social simulacra apart from real community behavior.

    “In evaluations, we show that participants are often unable to distinguish social simulacra from actual community behavior and that social computing designers successfully refne their social computing designs when using social simulacra.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    Social simulacra allow designers to explore hypothetical scenarios in which community members or moderators intervene.

    “and that they enable exploration of”
    • Also stated in: § 1 INTRODUCTION · plain
      “Likewise, social simulacra allow a designer to explore”
    Quote in the paperSubstantial claim
  • May combine two claims. 25
  • Classified as probably not a substantial claim. 26
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingreasoning

    Social simulacra prompted designers to revise their design by addressing more important edge cases in rules and by better defining and communicating cultural norms in the community goal.

    “This inspired them to iterate on their design by covering more important edge cases in their rules, as well as better scoping and communicating the cultural norms in their community goal statement.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the technical evaluation, participants misidentified an average of 41% (std=10) of pairs, which is close to chance performance.

    “Participants performed nearly at chance accuracy, misidentifying on average 41% (std=10) of pairs, suggesting that social simulacra can create plausible content.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the SimReddit without community description condition, participants' error rate was 21% (std=15%) and its pairwise comparison with SimReddit was significant (p < 0.01).

    “In the SimReddit without community description condition, our participants' error rate was 21% (std=15%) and its pairwise comparison between the SimReddit condition was signifcant ( < 0.01).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    In SimReddit, designers supply only a small set of example personas (10 by default), and the system expands them into many new personas (1,000 by default) that are distinct yet thematically related.

    “With SimReddit, our designers need to provide only a handful (10, by default) of example personas, and SimReddit uses these seed personas to generate a large number of new personas (1,000, by default) that are non-repeating but thematically relevant.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Creating hundreds of personas by hand can require substantial effort.

  • Classified as probably not a substantial claim. May combine two claims.

    Findingown data

    There was concern that iterating goals and rules of a live community would erode trust and make subsequent design changes less effective.

    “Furthermore, there was additional concern that iterating the goals and rules of a live community would erode the community's trust in the community management and turn subsequent design changes less efective”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingown data

    A persona consists of a name and a short description, and it shapes the topics and behaviors of individual generated users.

    “Personas contain a name and a simple descriptive phrase (e.g.,”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    SimReddit does not make every generated user follow every rule exactly; instead, rules act as nudges that push generated behavior toward the designer's intended direction.

    “Like in the real world, SimReddit does not enforce that all generated users follow all rules precisely.Instead, these rules are better interpreted as nudges that would encourage the generated behaviors to trend in the intended direction of the designer.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Rules alone can never stop determined trolls.

  • May combine two claims.

    Findingown data

    An omnibus ANOVA test confirms that these error rates are significantly different from each other (F(3, 196) = 22.49, p < 0.001), and posthoc Tukey tests confirmed that SimReddit significantly outperformed all other conditions.

    “An omnibus ANOVA test confrms that these error rates are signifcantly different from each other ( (3, 196) = 22.49, < 0.001), and posthoc Tukey tests between conditions confrmed that SimReddit signifcantly outperformed all other conditions (Table 1).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the SimReddit without personas condition, participants' error rate was 33% (std=10%) and its pairwise comparison with SimReddit was significant (p < 0.01).

    “And in the SimReddit without personas condition, our participants' error rate was 33% (std=10%) and its pairwise comparison between the SimReddit condition was signifcant ( < 0.01).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Some participants (n=3) recalled that rules are typically set reactively in response to problems.

    “In the absence of no better alternatives, however, some (n=3) recalled from their prior experience that”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    After a design is submitted, SimReddit produces a populated interface resembling a subreddit page, with content that reflects the given design specifications.

    “Once a design is submitted, SimReddit returns a populated interface that resembles a subreddit page (shown in Figure 1).The content on the page embodies the design specifcations provided.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    Some generated content was negative and prompted reflection on rules and moderation.

    “Meanwhile, some content was negative and prompted refection on rules and moderation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Participants did not believe the generations predicted what would happen; rather, they highlighted the tool's ability to ground their assumptions about the community.

    “But this was not because they believed that the generations predicted what is going to happen.Rather, they highlighted the tool's ability to ground their assumptions about the community.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    SimReddit performed significantly better than the crowdworker baseline and every ablation condition (p < .01).

    “SimReddit signifcantly outperformed the crowdworker baseline and all ablations ( < .01).”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    One-way ANOVA showed a significant effect of condition on error rate in the Technical Evaluation, F(3,196)=22.49, p<0.001.

    “Results from one-way ANOVA of participants' error rate in the Technical Evaluation, followed by Tukey's HSD post hoc test.*** p < 0.001; ** p < 0.01; * p < 0.05”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Anti-social behaviors like trolling, hate speech, inflammatory comments, and other undesirable acts can occur in social computing systems, making designs ineffective and causing harm to individuals and groups.

    “Anti-social behaviors such as trolling [26,38], hate speech [15], infammatory comments [12], and other”
    Quote in the paperSubstantial claim0 of 5 sources checked
    • cites [26] 26. Claire Hardaker. 2010. Trolling in asynchronous computer-mediated commu- nication: From user discussions to academic defnitions. Journal of Politeness Research 6 (2010), 215J242. source

      Not checked in the free run.

    • cites [38] 38. Joseph M Kayany. 1998. Contexts of uninhibited online behavior: Flaming in social newsgroups on Usenet. Journal of the American Society for Information Science 49, 12 (1998), 1135-1141.

      Not checked in the free run.

    • cites [45] 45. Robert E Kraut and Paul Resnick. 2012. Building successful online communities: Evidence-based social design. MIT Press.

      Not checked in the free run.

    • cites [2] 2. Yavuz Akbulut, Yusuf Levent Sahin, and Bahadir Eristi. 2010. Cyberbullying victimization among Turkish online social utility members. Journal of Educational Technology & Society 13, 4 (2010), 192-201.

      Not checked in the free run.

    • cites [77] 77. Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier bench- mark for general-purpose language understanding systems. Advances in neural J.S. Park, L. Popowski, C.J. Cai, M. Morris, P. Liang, M.S. Bernstein information processing systems 32 (2019).

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Instead of predicting one specific outcome, social simulacra can reveal a broader set of possible outcomes and let designers see how design changes might alter them.

    “Rather than making a single point prediction, social simulacra can surface a larger space of possible outcomes and enable the designer to explore how design changes might shift them.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Unless they actually launch systems at scale, designers currently lack any way to begin exploring these questions and reflecting on their designs' social dynamics.

  • May combine two claims.

    Interpretationreasoning

    Social simulacra let designers examine many possible outcomes and try out intervention strategies.

    “Social simulacra provide opportunities beyond generated behavior: they can also enable the designer to understand a multiverse of possible outcomes and to test intervention strategies.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Unless they actually launch systems at scale, designers currently lack any way to begin exploring these questions and reflecting on their designs' social dynamics.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    No model, now or in the future, can fully capture the subtlety and complexity of human behavior.

    “No model, present or future, can perfectly capture the nuance and complexity of human behaviors [66].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [66] 66. Valkyrie Savage, Colin Chang, and Björn Hartmann. 2013. Sauron: embedded single-camera sensing of printed physical user interfaces. In Proceedings of the 26th annual ACM symposium on User interface software and technology. 447-456.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The model is not aware of COVID-19 but infers from the phrase 'COVID vaccination' in the community description that COVID is a virus requiring vaccination.

    “Though the model is not aware of COVID-19, it infers from the phrase”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The study recruited 16 participants with prior experience designing or moderating online social spaces.

    “We recruited 16 participants who had prior experience designing or moderating online social spaces through social media and mailing lists.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Social simulacra is a general technique for leveraging generative models to prototype new social spaces.

    “Social simulacra is a general technique for leveraging generative models to prototype new social spaces.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The social simulacra technique can be generalized to prototype any similar social space beyond Reddit.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The only difference between the four study conditions was the way the synthetic conversations were gathered.

    “The only diference between these conditions was the way the synthetic conversations were gathered.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Large language models can perform many NLP tasks effectively-including language understanding, text classification, and generation-without fine-tuning.

    “Large language models can efectively perform a wide range of NLP tasks [5] such as language understanding [6,32,60,78], text classifcation [36,50,86], and generation [59,68] without the need for fne-tuning.”
    Quote in the paperSubstantial claim0 of 8 sources checked
    • cites [5] 5. Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021).

      Not checked in the free run.

    • cites [6] 6. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877-1901.

      Not checked in the free run.

    • cites [32] 32. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language under- standing. arXiv preprint arXiv:2009.03300 (2020).

      Not checked in the free run.

    • cites [60] 60. Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hofmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446 (2021).

      Not checked in the free run.

    • cites [78] 78. Laura Weidinger, John Mellor, Maribeth Rauh, Conor Grifn, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, et al. 2021. Ethical and social risks of harm from Language Models. arXiv preprint arXiv:2112.04359 (2021).

      Not checked in the free run.

    • cites [36] 36. Ellen Jiang, Kristen Olson, Edwin Toh, Alejandra Molina, Aaron Donsbach, Michael Terry, and Carrie J Cai. 2022. PromptMaker: Prompt-based Prototyping with Large Language Models. In CHI '22 Extended Abstracts on Human Factors in Computing Systems. source

      Not checked in the free run.

    • cites [50] 50. Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2021. Fantastically ordered prompts and where to fnd them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786 (2021). source

      Not checked in the free run.

    • cites [86]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    In the implementation, the GPT-3 prompt length restriction occasionally means earlier parts of a conversation must be truncated, and a broader set of posts to convey overall norms could not be fed to the model.

    “In our implementation, this restriction occasionally means we must truncate earlier parts of a conversation, and that we could not feed the model a broader set of posts to convey overall norms in the space.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    Real conversations were selected from threads at least 200 posts behind the latest and not in the top posts feed to avoid participants seeing them.

    “We searched for threads that were at least 200 posts behind the latest and did not appear in the top posts feed so that our participants would not come across the conversations used in the study to verify that the conversation actually took place.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Our prompting techniques boosted performance and made generations more plausible.

    “Our prompting techniques boosted performance and made generations more plausible.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    GPT-3 takes a natural language prompt as input and outputs a completion.

    “GPT-3 takes a natural language prompt as an input and outputs a completion.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Reactive strategies require harm to users and communities to improve; prototyping tools like social simulacra can minimize the need for these strategies.

    “Such reactive strategies thus require harm to users and communities to improve; prototyping tools like ours can minimize the need for these strategies.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Social simulacra carry a new risk: they may generate behaviors so far removed from reality that they fail to inform the designer's questions.

    “However, they introduce a new risk of possibly generating behaviors so detached from what might happen that they are not meaningful to the questions the designer wants to explore.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Careful prompt design can more effectively bring out the intended behavior from a model.

    “Furthermore, prompt design can better elicit the desired model behavior [36,37,49].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [36] 36. Ellen Jiang, Kristen Olson, Edwin Toh, Alejandra Molina, Aaron Donsbach, Michael Terry, and Carrie J Cai. 2022. PromptMaker: Prompt-based Prototyping with Large Language Models. In CHI '22 Extended Abstracts on Human Factors in Computing Systems. source

      Not checked in the free run.

    • cites [37] 37. Ellen Jiang, Edwin Toh, Alejandra Molina, Aaron Donsbach, Carrie J Cai, and Michael Terry. 2021. Genline and genform: Two tools for interacting with gen- erative language models in a code editor. In The Adjunct Publication of the 34th Annual ACM Symposium on User Interface Software and Technology. 145-147. source

      Not checked in the free run.

    • cites [49] 49. Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Gra- ham Neubig. 2021. Pre-train, prompt, and predict: A systematic survey of prompt- ing methods in natural language processing. arXiv preprint arXiv:2107.13586 (2021).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Large language models such as GPT-3 are a form of generative model with enough richness to support the assumptions that they can generate content in the relevant modality and encode enough world knowledge.

    “Large language models such as GPT-3 are one form of generative model encoding enough richness to support these assumptions.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Participants provided consent via an IRB-approved consent form.

    “The participants provided consent by agreeing to a consent form that was approved by our institution's IRB.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Even with a multiverse, social simulacra are unlikely to present an exhaustive list of possible outcomes in a social system.

    “Even with a multiverse, social simulacra are unlikely to present an exhaustive list of possible outcomes in a social system.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Large language models trained on a user's chat history can predict that user's future replies with some accuracy, produce realistic action plans, or draw out varied commonsense reasoning.

    “So, large language models trained on a user's chat history can predict the user's future responses with some accuracy [47,79], generate realistic action plans that an embodied agent might take [34], or elicit a large variety of commonsense reasoning [48].”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [47] 47. Mike Lewis, Denis Yarats, Yann N Dauphin, Devi Parikh, and Dhruv Batra. 2017. Deal or no deal? end-to-end learning for negotiation dialogues. arXiv preprint arXiv:1706.05125 (2017). source

      Not checked in the free run.

    • cites [79] 79. Tongshuang Wu, Ellen Jiang, Aaron Donsbach, Jef Gray, Alejandra Molina, Michael Terry, and Carrie J Cai. 2022. PromptChainer: Chaining Large Language Model Prompts through Visual Programming. arXiv preprint arXiv:2203.06566 (2022).

      Not checked in the free run.

    • cites [34] 34. Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. 2022. Lan- guage Models as Zero-Shot Planners: Extracting Actionable Knowledge for Em- bodied Agents. arXiv preprint arXiv:2201.07207 (2022).

      Not checked in the free run.

    • cites [48] 48. Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi, and Hannaneh Hajishirzi. 2021. Generated knowledge prompting for commonsense reasoning. arXiv preprint arXiv:2110.08387 (2021).

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Participants were randomly assigned into one of four study conditions.

    “We randomly assigned our participants into one of the four study conditions.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    Social simulacra use large language models to fill a social computing system with believable social behaviors.

    “Social simulacra leverage large language models to populate a social computing system with plausible social behaviors.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Designers are often caught off guard by the behaviors that appear once their spaces are fully populated.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    One implication of this work is a turn towards proactive, instead of reactive, design for social computing systems.

    “One implication of this work is a turn towards proactive, instead of reactive, design for social computing systems.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Populating social computing systems is often a necessary but nearly impossible challenge, since emergent behaviors at cold start differ from those at critical mass.

    “But populating social computing systems, particularly for those that are designed to support interactions of a large number of people, is often an insurmountable, though necessary, challenge because the emergent behaviors at cold start are diferent than those when the space hits critical mass [4].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [4] 4. Michael S. Bernstein, Mark S. Ackerman, Ed H. Chi, and Robert C. Miller. 2011. The Trouble with Social Computing Systems Research. In CHI '11 Extended Abstracts on Human Factors in Computing Systems (Vancouver, BC, Canada) (CHI EA '11). Association for Computing Machinery, New York, NY, USA, 389-398. https://doi.org/10.1145/1979742.1979618

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Having many varied personas is important for revealing how different personas might interact and what different intentions they bring to the space.

    “A large and diverse set of personas is important for surfacing how diferent personas might interact with each other, as well as their difering intentions within the space.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Social simulacra fulfill the role of early prototyping techniques by pushing designers to question assumptions and consider a broader design space.

    “Social simulacra fulfll the role that many of the early prototyping techniques fulfll outside of social computing design: they push the designer to question their assumptions, and through this, inspire them to consider a broader design space.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Subreddit designers must set the community goal and rules and decide implicit policies like target population and moderation strategies; these choices help shape social interactions in the community.

    “They need to defne the community goal (e.g.,”
    • Also stated in: § 3.1 System Context · plain
      “These elements contribute to shaping the social interactions in the community [11,17,53].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [11] 11. Eshwar Chandrasekharan, Mattia Samory, Shagun Jhaver, Hunter Charvat, Amy Bruckman, Clif Lampe, Jacob Eisenstein, and Eric Gilbert. 2018. The Internet's Hidden Rules: An Empirical Study of Reddit Norm Violations at Micro, Meso, and Macro Scales. Proc. ACM Hum.-Comput. Interact. 2, CSCW, Article 32 (nov 2018), 25 pages. https://doi.org/10.1145/3274301 source

      Not checked in the free run.

    • cites [17] 17. Casey Fiesler, Joshua McCann, Kyle Frye, Jed R Brubaker, et al. 2018. Reddit rules! characterizing an ecosystem of governance. In Twelfth International AAAI Conference on Web and Social Media. source

      Not checked in the free run.

    • cites [53] 53. Nathan J. Matias. 2019. Preventing harassment and increasing group participation through social norms in 2,190 online science discussions. Proceedings of the National Academy of Sciences 116, 20 (2019), 9785-9789. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Social simulacra cannot substitute for actual participatory engagement with users.

    “Social simulacra cannot and should not replace actual participatory engagement with users.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductreasoning

    In our prototypes, we reduced the risk of bias replication by using a broad range of seed personas.

    “In our prototypes, we mitigated this risk by ensuring a broad range of seed personas; this could become a more explicit test in future versions of our system.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    This bias replication is problematic if it leads designers to overlook the needs of these groups in design.

    “This is problematic if it causes designers to then overlook these peoples' needs in design.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    There is a risk that designers may over-rely on simulacra, leading to downstream issues.

    “There is a risk that designers may overrely on simulacra, which can lead to issues downstream.”
    Quote in the paperNot a substantial claim
46 of 46

  • The paper relies on this without support.

    Designers are often caught off guard by the behaviors that appear once their spaces are fully populated.

    “Yet, despite decades of progress in research and practice, understanding the efects of these design decisions remains challenging; as a result, designers are regularly surprised by the behaviors that arise when their spaces are fully populated.”

  • The paper relies on this without support.

    Unless they actually launch systems at scale, designers currently lack any way to begin exploring these questions and reflecting on their designs' social dynamics.

    “Barring actually launching our systems at scale, designers currently have no way of starting to explore these questions to refect on the social dynamics of their designs.”

  • The paper relies on this without support.

    Creating hundreds of personas by hand can require substantial effort.

    “However, manually crafting hundreds of personas can be labor intensive.”

  • The paper relies on this without support.

    Rules alone can never stop determined trolls.

    “There are still some trollish behaviors left, but Sam comes to recognize that rules are never enough to stop dedicated trolls and that she will have to moderate the community.”

  • The authors state this as an assumption.

    Social simulacra assume the model can generate content in the modality relevant to the design space.

    “Social simulacra make two assumptions about the model used for their generations: frst, the model needs to be able to generate content in the modality relevant to the design space of the system (e.g., text)”

  • The authors state this as an assumption.

    Social simulacra assume the model encodes enough knowledge about the world and people to generate content relevant to the designer's questions.

    “second, it must encode enough knowledge about the world and people so that it can generate content relevant to the design questions that the designer wants to answer.”

  • The paper relies on this without support.

    Subreddit threads are structured with top-level posts that introduce a topic and replies that engage in discussion while adhering to community rules, and participants maintain consistent yet diverse personas.

    “But, subreddit threads are structured with toplevel posts that introduce a topic relevant to the community, and replies engage in discussion on the topic while (typically) adhering to the community rules.Moreover, participants need to maintain consistent yet diverse personas.”

  • The paper relies on this without support.

    Current large language models such as GPT-3 have a strict character limit to their input.

    “Current large language models such as GPT-3 have a strict character limit to their input.”

  • The paper relies on this without support.

    GPT-3 can produce multiple different outputs given one input prompt.

    “GPT-3 can produce multiple diferent outputs given one input prompt.”

  • The paper relies on this without support.

    A generation requires a large collective of personas to ensure diversity in its content.

    “but a generation requires a large collective of personas to ensure diversity in its content.”

  • The paper relies on this without support.

    Social computing designers often prototype in unexplored contexts.

    “We often prototype in unexplored contexts.”

  • The paper relies on this without support.

    Embedding these characteristics in generated content is critical to creating useful design tools.

    “Embedding these characteristics in the generated content is critical to creating useful design tools.”

  • The paper relies on this without support.

    The realism of generated content is sufficient for it to be useful in informing design decisions.

    “There is enough realism in them to make them compelling and we can easily imagine that they might happen in the future.”

  • The paper relies on this without support.

    Design revisions made in response to simulated interactions constitute improvements to the community.

    “The initial generations prompted all but one of our 16 participants to revise their original design.”

  • The paper relies on this without support.

    The social simulacra technique can be generalized to prototype any similar social space beyond Reddit.

    “Social simulacra is a general technique for leveraging generative models to prototype new social spaces.”

  • The paper relies on this without support.

    Women and minoritized groups are frequently silenced in online spaces.

    “Women and minoritized groups, for example, are often silenced in online spaces”

  • The paper relies on this without support.

    Models like GPT-3 may have learned these silencing patterns and reproduce them.

    “models such as GPT-3 may have learned these patterns and replicate them”
17 of 17

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Park et al. (2022) show that LLM-based agents are capable of generating social media posts indistinguishable from those written by humans.

    “Additionally, Park et al. (2022) show that LLM-based agents are capable of generating social media posts indistinguishable from those written by humans.”

    From Simulating Opinion Dynamics with Networks of LLM-based Agents (2024) · cites this paper as Park et al., 2022

    The Study Procedure describes a test in which participants distinguished SimReddit conversations from real subreddit conversations. In Results (§5.5; page unavailable), participants misidentified them 41% of the time—close to, but not the same as, the 50% expected from random guessing. The study supports that LLM-generated social media conversations can be difficult to distinguish from human ones, but not the unqualified claim that the posts are indistinguishable. (LLM, conf=0.88)

    Experimental — accuracy not yet measured

  • The social behaviors observed in Smallville arise emergently rather than being pre-programmed.

    “Extending prior work [80], these social behaviors are emergent rather than pre-programmed.”

    From Generative Agents: Interactive Simulacra of Human Behavior (2023) · cites this paper as [80]

    Section 3 describes SimReddit generating users and their interactions with GPT-3, supporting the prior-work premise that social behaviors can be generated rather than individually pre-programmed. The paper does not examine Smallville, so it cannot establish how Smallville’s particular behaviors arose. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • The condition lacking observations, reflections, and planning effectively represents the prior state of the art for large-language-model agents.

    “The no observation, no refection, no planning condition efectively represents the previous state of the art for agents created through large language models [12,46,80].”

    From Generative Agents: Interactive Simulacra of Human Behavior (2023) · cites this paper as [80]

    Section 4.2 describes a prior GPT-3 system that generates social participants through persona-based prompt chains, so this paper is an example of an earlier language-model agent approach. But §4.2.3 says replies are prompted with the prior conversation as context; the paper does not identify its system as a no-observation, no-reflection, no-planning condition or establish that condition as the previous state of the art. (LLM, conf=0.78)

    Experimental — accuracy not yet measured

  • Social evaluation assesses social intelligence based on agent interactions in simulated societies, using approaches such as collaborative tasks, debates, and human studies.

    “This method utilizes metrics to assess social intelligence based on the agent interactions in simulated societies.Various approaches have been adopted, such as collaborative tasks to evaluate teamwork skills, debates to analyze argumentative reasoning, and human studies to measure social aptitude [34,98,102,165,166].”

    From A survey on large language model based autonomous agents (2024) · cites this paper as [98]

    The Abstract describes simulated interactions among community members, and §5.2 describes a human study comparing generated conversations with real ones (page numbers unavailable). But §5 says the evaluations test the realism of those conversations and their usefulness to designers, not agents’ social intelligence. The paper does not support the attributed examples of collaborative tasks measuring teamwork, debates measuring reasoning, or human studies measuring social aptitude. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • Park et al. (2022) use foundation models without safety filters because their research aims to simulate human behavior including toxic speech, and most closed foundation models would suppress these outputs.

    “Park et al. (2022) use foundation models without safety filters because their research aims to simulate human behavior (including toxic speech).Most closed foundation models would suppress these outputs.”

    From On the Societal Impact of Open Foundation Models (2024) · cites this paper as Park et al. (2022)

    Section 2.3 says the authors use GPT-3 to simulate social behavior and regard its ability to generate troll behavior as useful for design. Section 4.3 identifies the base davinci engine and its settings, but does not say they chose a model without safety filters for that reason. The paper also does not say that most closed foundation models would suppress toxic outputs. (LLM, conf=0.87)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Crowdworker gender counts exceed the reported sample size

    Minor

    The reported counts of 28 female and 32 male crowdworkers total 60, although the paper says it recruited 50 crowdworkers.

    Reported 50 · recomputed 60

  • Reply example changes the post it says it continues

    Minor

    The reply-generation example is introduced as a reply to Layla Li's post above, but its thread contains a different Layla Li post from the one generated in the preceding example.

  • Ablation error rate differs between results and table

    Minor

    The results text gives a 33% error rate without personas, while Table 1 gives 34% for that same condition.

3 of 3

The citation lists haven’t been fetched yet

References 57 · Cited by 212

Research works printed but not linked: 17

  1. Tom Brown , Benjamin Mann , Nick Ryder , Melanie Subbiah , Jared  D Kaplan , Prafulla Dhariwal , Arvind Neelakantan , Pranav Shyam , Girish Sastry , Amanda Askell , 2020. Language models are few-shot learners. Advances in neural information processing systems 33 ( 2020 ), 1877–1901. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877–1901.
  2. Justin Cheng , Cristian Danescu-Niculescu-Mizil , and Jure Leskovec . 2015 . Antisocial behavior in online discussion communities . In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 9. 61–70 . Justin Cheng, Cristian Danescu-Niculescu-Mizil, and Jure Leskovec. 2015. Antisocial behavior in online discussion communities. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 9. 61–70.
  3. Joan Donovan . 2019. How hate groups ’ secret sound system works. The Atlantic 17( 2019 ). https://doi.org/10. 1145/1188913.1188915 Joan Donovan. 2019. How hate groups’ secret sound system works. The Atlantic 17(2019). https://doi.org/10. 1145/1188913.1188915
  4. Bill Gaver , Tony Dunne , and Elena Pacenti . 1999. Design: cultural probes. interactions 6, 1 ( 1999 ), 21–29. Bill Gaver, Tony Dunne, and Elena Pacenti. 1999. Design: cultural probes. interactions 6, 1 (1999), 21–29.
  5. 10.1145/1449715.1449732 · DOI
  6. Stephanie Houde and Charles Hill . 1997. What do prototypes prototype?In Handbook of human-computer interaction ( second ed.). Elsevier , 367–381. Stephanie Houde and Charles Hill. 1997. What do prototypes prototype?In Handbook of human-computer interaction(second ed.). Elsevier, 367–381.
  7. M.  Lynne Markus . 1987 . Toward a “Critical Mass ” Theory of Interactive Media. Communication Research 14 (1987), 491 – 511 . M. Lynne Markus. 1987. Toward a “Critical Mass” Theory of Interactive Media. Communication Research 14 (1987), 491 – 511.
  8. Prolific. 2022 . Prolific: Quickly find research participants you can trust. https://www.prolific.co/ Prolific. 2022. Prolific: Quickly find research participants you can trust. https://www.prolific.co/
  9. Alec Radford , Jeffrey Wu , Rewon Child , David Luan , Dario Amodei , Ilya Sutskever , 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 ( 2019 ), 9. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9.
  10. Horst  WJ Rittel and Melvin  M Webber . 1973. Dilemmas in a general theory of planning. Policy sciences 4, 2 ( 1973 ), 155–169. Horst WJ Rittel and Melvin M Webber. 1973. Dilemmas in a general theory of planning. Policy sciences 4, 2 (1973), 155–169.
  11. Carolyn Snyder. 2003. Paper prototyping: The fast and easy way to design and refine user interfaces. Morgan Kaufmann. Carolyn Snyder. 2003. Paper prototyping: The fast and easy way to design and refine user interfaces. Morgan Kaufmann.
  12. David Thomas . 2006. A General Inductive Approach for Analyzing Qualitative Evaluation Data. American Journal of Evaluation27 ( 2006 ). David Thomas. 2006. A General Inductive Approach for Analyzing Qualitative Evaluation Data. American Journal of Evaluation27 (2006).
  13. Emily  A Vogels . 2021. The state of online harassment . Pew Research Center 13( 2021 ). Emily A Vogels. 2021. The state of online harassment. Pew Research Center 13(2021).
  14. Alex Wang , Yada Pruksachatkun , Nikita Nangia , Amanpreet Singh , Julian Michael , Felix Hill , Omer Levy , and Samuel Bowman . 2019 . Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems 32 (2019). Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems 32 (2019).
  15. Tongshuang Wu Ellen Jiang Aaron Donsbach Jeff Gray Alejandra Molina Michael Terry and Carrie J Cai. 2022. PromptChainer: Chaining Large Language Model Prompts through Visual Programming. arXiv preprint arXiv:2203.06566(2022). Tongshuang Wu Ellen Jiang Aaron Donsbach Jeff Gray Alejandra Molina Michael Terry and Carrie J Cai. 2022. PromptChainer: Chaining Large Language Model Prompts through Visual Programming. arXiv preprint arXiv:2203.06566(2022). · arXiv
  16. Rowan Zellers , Ari Holtzman , Hannah Rashkin , Yonatan Bisk , Ali Farhadi , Franziska Roesner , and Yejin Choi . 2019. Defending against neural fake news. Advances in neural information processing systems 32 ( 2019 ). Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. Advances in neural information processing systems 32 (2019).
  17. Zihao Zhao , Eric Wallace , Shi Feng , Dan Klein , and Sameer Singh . 2021 . Calibrate before use: Improving few-shot performance of language models . In International Conference on Machine Learning. PMLR, PMLR, 12697–12706 . Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning. PMLR, PMLR, 12697–12706.

Authors
Park, J. S., Popowski, L., Cai, C., Morris, M. R., Liang, P., & Bernstein, M. S.
Affiliations
  • Stanford University (US)
  • Google (United States) (US)
ORCID
Venue
Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology
Publisher
ACM
Published
2022-10-28
Pages
1-18
DOI
10.1145/3526113.3545616
OpenAlex
W4307475457
Cited by
212
References
57
Citations by year
2026: 45 · 2025: 80 · 2024: 58 · 2023: 29
FWCI
33.41
Type
conference paper
Language
en
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor)
Locations
  • ACM Symposium on User Interface Software and Technology (UIST) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Topics
  • Physical Sciences › Computer Science › Human-Computer Interaction › Innovative Human-Technology Interaction
  • Physical Sciences › Computer Science › Human-Computer Interaction › Usability and User Interface Design
  • Physical Sciences › Computer Science › Information Systems › Recommender Systems and Techniques
Keywords
large language models, antisocial behavior, social media communities, personas, what-if scenarios, community moderation, design refinement

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-09-28.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-09-28.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review