Simulation as Alibi---How the Social Order Gets Outsourced to LLM Agents

Abstract

LLM agent societies are moving rapidly from academic sandboxes into commercial platforms, military infrastructure, and government operations. Across these domains, the language of emergence—the claim that coordination patterns, norms, and consensus arise spontaneously from local interaction—has become the default vocabulary for describing multi-agent behavior. This paper argues that emergence discourse performs institutional work beyond scientific description: it renders invisible the procedural architecture—interaction topologies, action grammars, sanction architectures, and beneficiary structures—authored by identifiable actors. Drawing on the concept of the alibi from critical humanitarian studies, we develop the simulation as alibi framework, comprising three mechanisms: epistemic laundering, accountability displacement, and legitimacy extraction. Through critical discourse analysis of 20 research papers and 5 platform documents, systematic coding of 142 papers on LLM agent societies, institutional gap analysis of 6 governance frameworks, and cross-domain case studies spanning commercial, military, and governmental settings, we demonstrate three convergent findings. First, in our corpus, claims of emergence are almost always associated with procedurally conditioned architectures. Second, the governance frameworks examined lack the conceptual resources to detect this condition. Third, the resulting accountability gap is distributionally stable because the actors positioned to close it benefit from its persistence. We propose Upstream Disclosure Obligations for Agent Societies (UDOS), which extend accountability from deployment-stage auditing to the upstream design stage where procedural architecture is authored. UDOS shifts the regulatory question from what does this model do? to who wrote the rules under which this model interacts?