From Plausible to Causal: Counterfactual Semantics for Policy Evaluation in Simulated Online Communities Conference: PoliSim@CHI 2026: LLM Agent Simulation for Policy; April 16, 2026; Barcelona, Spain CCS: Human-centered computing Empirical studies in HCI CCS: Human-centered computing Social media
Agam Goyal, Yian Wang, Eshwar Chandrasekharan, Hari Sundaram · 2026
Read From Plausible to Causal: Counterfactual Semantics for Policy Evaluation in Simulated Online Communities Conference: PoliSim@CHI 2026: LLM Agent Simulation for Policy; April 16, 2026; Barcelona, Spain CCS: Human-centered computing Empirical studies in HCI CCS: Human-centered computing Social media by Agam Goyal, Yian Wang, Eshwar Chandrasekharan, Hari Sundaram (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-03920.txt.
Introduction
Agam Goyal and colleagues argued that lifelike mock communities do not yet prove a policy works. Their frame splits two kinds of cause. Necessary causation means the result would not have come about without the move. Sufficient causation means the move brings the result about time and again. Staff who probe past events need the first kind, while staff who pick new rules need the second. All such claims still hinge on how true the mock world is, so lifelike looks alone cannot back a rule change.
Catalog record
| CREATOR | Agam Goyal, Yian Wang, Eshwar Chandrasekharan, Hari Sundaram |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27376 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.03920
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
From Plausible to Causal: Counterfactual Semantics for Policy Evaluation in Simulated Online Communities Conference: PoliSim@CHI 2026: LLM Agent Simulation for Policy; April 16, 2026; Barcelona, Spain CCS: Human-centered computing Empirical studies in HCI CCS: Human-centered computing Social media Agam Goyal, Yian Wang, Eshwar Chandrasekharan, Hari Sundaram Abstract . LLM-based social simulations can generate believable community interactions, enabling “policy wind tunnels” where governance interventions are tested before deployment. But believability is not causality. Claims like “intervention A A reduces escalation” require causal semantics that current simulation work typically does not specify. We propose adopting the causal counterfactual framework, distinguishing necessary causation (would the outcome have occurred without the intervention?) from sufficient causation (does the intervention reliably produce the outcome?). This distinction maps onto different stakeholder needs: moderators diagnosing incidents require evidence about necessity, while platform designers choosing policies require evidence about sufficiency. We formalize this mapping, show how simulation design can support estimation under explicit assumptions, and argue that the resulting quantities should be interpreted as simulator-conditional causal estimates whose policy relevance depends on simulator fidelity. Establishing this framework now is essential: it helps define what adequate fidelity means and moves the field from simulations that look realistic toward simulations that can support policy changes. 1. The Causal Gap in Simulation-Based Governance of Online Communities Online communities today are not merely social spaces, but rather they are governance systems ( 34 ) . Every moderation action or rule enactment on platforms such as Reddit and X constitutes a policy decision with consequences for millions of users ( 20 ) . These policy decisions have widespread consequences such as community norm ( 5 ) , linguistic dynamics ( 11 ; 2 ) , newcomer retention ( 31 ) , and public perception of the platform. Despite these stakes, community governance remains largely reactive ( 32 ) , where platform designers and moderators look at outcomes of policy changes and interventions in a post-hoc manner which forecloses a huge opportunity for proactive governance which can help in anticipating how interventions might reshape community dynamics before deployment ( 22 ; 23 ) . The HCI research community has previously identified the lack of prototyping techniques in social computing as a key reason for this gap for this gap ( 28 ) , highlighting that designers lack robust ways to proactively reflect on the downstream consequences that may emerge from such interventions ( 33 ) . As a result, the paradigm of LLM-based social simulations has emerged as a lightweight prototyping technique to test design choices and their potential downstream consequences ( 27 ) . However, in this paper we argue that currently, these simulation-based prototyping techniques can help us map a set of plausible outcomes, but cannot yet answer causal questions—whether deployment of intervention A will reliably cause outcome B. To see why this distinction matters, consider the kinds of questions that governance stakeholders need to answer. A moderator reviewing a thread that escalated to a toxic outcome would ask “Was the initial inflammatory comment the cause of this outcome? Would the thread have stayed civil without it?” A platform designer evaluating an automated counter-speech system would ask “If we deploy this at scale, will it reliably prevent escalation?” Finally, a policy team weighing rule changes would ask “How much of the toxicity reduction we observed was due to our intervention versus an independent shift in community norms?” These are causal questions, and none of them can be answered by demonstrating that a simulation produces believable content. We argue that, for simulations to inform governance decisions, the field must move from asking “Do these simulations look real?” to asking “What causal claims can these simulations support?” This is not a critique of existing work: progress on behavioral plausibility has been essential and foundational. Rather, we see this as the next conceptual step. Governance stakeholders do not only need evidence that an intervention changes outcomes on average; they need to know whether a specific factor was necessary for a harmful outcome, whether an intervention is sufficient to reliably prevent it, and under what assumptions such conclusions hold. If simulations are to inform these decisions, the evidence they produce must be legible in causal terms, for example, “the probability that the outcome would not have occurred without the intervention is X .” To this end, we propose adopting the causal counterfactual framework for event attribution developed by 16 to analyze simulated interventions in online communities. Its key advantage is that it distinguishes two different causal questions that governance stakeholders routinely conflate: whether a factor was necessary for a harmful outcome, and whether an intervention is sufficient to reliably produce a desired one. This distinction enables simulations to support precise, probabilistic, and policy-relevant claims from model outputs, rather than only informal judgments that an intervention “seemed to help.” The framework has been influential in fields such as climate-event attribution, and we argue that it provides the missing causal semantics for simulation-based policy evaluation in social computing. The remainder of this paper develops our central position. Section 2 presents an overview of the literature landscape on LLM simulations. Section 3 introduces the causal counterfactual framework, grounding its key concepts in concrete governance examples. Section 4 maps the framework onto simulation-based policy evaluation, showing how different types of causal evidence serve different governance stakeholders. Section 5 presents an important pre-requisite about simulator fidelity under this framework. Finally, Section 6 outlines a research agenda for operationalizing causal claims in simulated communities. 2. Related Works Recent work on LLM-based social simulation has demonstrated that language-model agents can generate plausible social interactions and support early-stage design exploration for social computing systems. For social computing specifically, 28 introduced the idea of populated prototypes, using language models to help designers reason about social interactions before deployment, and showed how generative agents can produce believable, persistent, and socially situated behavior in interactive environments ( 27 ) . More recent work has pushed this paradigm toward larger-scale and more empirically grounded simulation, including efforts to simulate populations and known human-group dynamics at scale ( 29 ; 35 ; 9 ; 8 ; 7 ) . These works overall argue that LLM social simulations may become a useful research method under appropriate validation and scope conditions ( 1 ) . Domain-specific work has also begun to study content dissemination and regulation in multi-agent social media simulation settings ( 25 ; 3 ) including agent-native social networks such as Moltbook ( 10 ; 13 ) . However, this literature has primarily focused on plausibility, human-behavioral resemblance, or methodological promise, rather than on the semantics of what kinds of causal claims simulated interventions can justify. At the same time, the online community governance literature has established that real interventions such as bans, quarantines, moderation explanations, and visible norm-setting can meaningfully reshape participation and discourse ( 6 ; 5 ; 17 ; 26 ; 4 ; 12 ; 34 ; 11 ) . This line of work provides both concrete intervention classes and real-world outcome measures against which simulation-based policy evaluation could eventually be calibrated. Our contribution is complementary to both strands, and our central position is that policy-facing simulation requires an explicit causal semantics. To that end, we adapt