AI Post Transformers · Visual Companion

Social Simulacra for Prototyping Online Communities

A visual wind tunnel for a synthetic online community: from a goal, a rulebook, and a few seed personas to generated members, reply trees, moderation branches, and the realism gap between human-sounding text and believable institutions.

Primary paper · arXiv 2208.04024 UIST 2022 LLM populated prototype Transcript IDs: 2208.04024
10 seed personas -> synthetic crowd 41% misidentification of simulated threads 15/16 designers revised rules 4 moderator moves replayed

1

community charter anchors the simulation before any crowd exists.

~1k

synthetic members are scaled by prompt chaining, not model retraining.

4

warning, removal, restatement, and redirect branch thread outcomes.

2

separate questions: local plausibility versus long-run social realism.

System Map

From community idea to social weather system

Step through the pipeline to see where the method adds scale and where it still depends on short, text-only counterfactual scenes.

Prompt-chained simulation pipeline

Boxes, flows, and risk rails. The selected stage expands below and the right-hand bars show which failure modes become visible there.

structure population synthesis risk surfacing

Synthetic Members

Seeds can stay distinct, or collapse into drift

Toggle the social regime and scrub the rounds. The heatmap tracks persona priors; the field plot shows whether the simulated crowd preserves diversity or converges into one hotter tone band.

Persona trait heatmap

Rows are seed personas. Columns are behavioral priors that feed generated posts and replies.

Population field

Cluster geometry, conflict temperature, and rule adherence after repeated interactions.

Moderation Replay

Textual interventions bend the thread, but do not create an institution

Pick a moderator move. The tree changes locally, while the outcome chart tracks on-topic behavior, harassment, retention, mod load, and clarity.

Reply tree after intervention

Green and cyan branches cool the conversation. Orange and red show piling on. Removed nodes are crossed out.

Outcome deltas

Gray is the baseline. The selected intervention overlays with a sparkline for conflict temperature across eight rounds.

Method Limits

Plausible threads are not the same thing as believable communities

Sort by design speed or realism. Social Simulacra lands high on speed and control, lower on real stakes, memory, and feedback loops.

Prototype method matrix

Heatmap of control, speed, stakes, memory, tail-risk coverage, feedback loops, and stakeholder coverage.

Local plausibility versus community realism

The episode's key critique in one plot: a thread can sound human before the incentive system feels real.

References

Papers and listening context

Compact source cards. Only one explicit arXiv ID appears in the transcript: 2208.04024.