Tang, S., & Lin, Z. (2026). Ninth AAAI/ACM Conference on Artificial Intelligence, Ethics and Society (AIES-26), camera-ready.
This repository is the audit and supporting-materials companion to the paper. It serves three purposes:
- A field overview: a curated, coded catalog of 142 papers on LLM agent societies (2022–2026), organized by emergent phenomenon.
- The paper's evidence: the full coding data behind Table 2 and every statistic reported in Section 4.2, reproducible with one command.
- The coding protocol: the emergence-explanation-sentence rubric behind the designer-visibility (DV) dimension.
The systematic coding corpus tracks the rapid expansion of LLM agent society research. Peer-reviewed papers (29) span 2023–2026 steadily; arXiv preprints (110) explode in 2025–2026, reaching 68 papers in 2026 alone. ~80% of the corpus are preprints — a deliberate coverage choice, because the central debates over emergent behavior are currently unfolding in the preprint literature, as discussed in the paper's limitation section. The preprint-heavy composition is itself informative: the strongest emergence claims — and the strongest instances of designer erasure — are concentrated precisely where discourse is least institutionally constrained.
What the paper argues. Across this corpus, claims of emergence are almost always associated with procedurally conditioned architectures: the coordination, norms, and "spontaneous" behaviors attributed to agent self-organization are conditioned by authored interaction topologies, action grammars, sanction architectures, and beneficiary structures. Yet the governance frameworks examined (EU AI Act, NIST AI RMF 1.0, DoD AI Acceleration Strategy, White House EO, EU Emergency Management Framework, China MOST Guidelines) contain no explicit architecture-level procedural authorship disclosure mechanism. The paper develops simulation as alibi — three mechanisms (epistemic laundering, accountability displacement, legitimacy extraction) through which emergence discourse obscures procedural authorship — and proposes Upstream Disclosure Obligations for Agent Societies (UDOS).
LLM social simulations do not simply reveal social order. Their architectures help produce it, but current AI governance rarely requires simulation designers to disclose those choices.
LLM social simulations now model public opinion, markets, and crisis response. Their outputs can inform high-stakes decisions, but governance frameworks do not yet capture procedural authorship.
Our audit of 142 papers finds that designer erasure concentrates where claims of emergence are strongest. Fourteen of the 17 papers with absent designer visibility fall in the High-EI group.
The Alibi Function identifies three mechanisms: epistemic laundering, accountability displacement, and legitimacy extraction. UDOS requires researchers to disclose the interaction architecture before outputs travel.
Download the editable figures or reproduce them from docs/social_cards.
All 142 papers, each with title, venue, year, arXiv link (where available), abstract, and the audit summary (EI, DV, designer-in-sentence, RT). Organized by emergent phenomenon:
| Category | Papers | File |
|---|---|---|
| Coordination & Cooperation Dynamics | 31 | 01_coordination.md |
| Cooperation & Social Dilemmas | 33 | 02_cooperation.md |
| Bias, Polarization & Inequality | 29 | 03_bias.md |
| Culture, Emotion & Normativity | 8 | 04_culture.md |
| Norms & Conventions | 9 | 05_norms.md |
| Opinion Dynamics & Social Influence | 8 | 06_opinion.md |
| Role Specialization & Networks | 3 | 07_roles.md |
| Collective Cognition & Intelligence | 5 | 08_cognition.md |
| Simulation Methods & Platforms | 8 | 09_methods.md |
| Other Emergent Phenomena | 8 | 10_other.md |
Reading an entry. Each paper entry reports: the verbatim emergence explanation sentence (the sentence in which the paper explains why the phenomenon occurs), and three audit dimensions — EI (emergence intensity: how strongly the paper asserts spontaneous emergence), DV (designer visibility: whether the design choices appear in the explanation sentence as a causal subject), and RT (regulatory translatability). The signature finding of the paper is visible across the catalog: papers with the strongest emergence claims are disproportionately the ones in which the designer has disappeared from the explanation sentence.
Each catalog entry is coded on the paper's audit dimensions. The definitions below provide the interpretation context for reading the catalog files.
- EI — Emergence intensity (
Low/Medium/High, coded 1/2/3): how strongly, and with how little qualification, a paper asserts that behavior arises spontaneously. High asserts emergence as an established fact ("norms spontaneously emerged"); Medium presents it as an interpretive finding with hedging ("our results suggest the emergence of…"); Low mentions emergence only as a secondary observation. - DV — Designer visibility (
Absent/Partial/Full, coded 1/2/3): how explicitly a paper attributes the interaction architecture to identifiable design choices, anchored on the emergence explanation sentence. Full names a design choice as the causal subject ("removing memory prevents the emergence of stable cooperation"); Partial discloses the architecture in the methods but the explanation sentence does not return to the designer; Absent narrates emergence as spontaneous, self-organized, or population-level ("norms emerged naturally from interactions"). Risk, governance, and ethical discussions do not count as evidence. - Designer in explanation sentence (
Yes/No): whether the designer or a design choice appears in the emergence explanation sentence as a causal subject. - DP — Deployment proximity (
1= research only,2= deployment implied): whether emergence is discussed only in a research context, or in a deployment/policy context (commercial applications, governance recommendations, policy citations). - RT — Regulatory translatability (
1= Low,2= Medium,3= High): how readily a paper's description of emergent behavior translates into governance obligations. High is directly translatable; Medium is partially translatable but missing key information; Low is too abstract to yield concrete governance requirements.
The emergence explanation sentence is the sentence in which a paper explains why an emergent phenomenon occurs (typically Results/Discussion); DV and Designer-in-sentence codes are anchored on it. EI and DV are coded independently on distinct criteria — EI from the assertion of spontaneity, DV from the attribution of authorship — so a paper may assert strong emergence while still crediting specific design choices.
python3 scripts/reproduce_statistics.pyReproduces every statistic in the paper from
data/stream1b_systematic_coding/coding_data_142.json:
| Statistic | Paper | Reproduced |
|---|---|---|
| χ² (df=4) | 16.62 | 16.622 |
| p | 0.002 | 0.002 |
| Cramér's V | 0.242 | 0.242 |
| r(EI, DV) | -0.141 (p = 0.094) | -0.141 |
| r(EI, RT) | 0.194 (p = 0.021) | 0.194 |
| Absent concentration | 14/17 (82%) in High EI | 14/17 (82%) |
| arXiv share | ~80% | 80% |
simulation-as-alibi/
├── README.md # this file
├── docs/
│ └── corpus_by_year.svg # corpus chart (peer-reviewed vs preprint)
├── data/
│ ├── stream1b_systematic_coding/
│ │ ├── catalog/ # paper-by-paper audit, 10 category files
│ │ ├── coding_data_142.json # machine-readable coding data
│ │ ├── cross_tabulation_ei_dv.csv
│ │ ├── rationales_worksheet.csv
│ │ └── README.md
│ ├── stream1a_cda/README.md # qualitative CDA stream (no data published)
│ ├── stream2_platforms/ # 5 platforms + case audit
│ └── stream3_governance/ # 6 governance frameworks + official sources
├── codebook/
│ └── codebook_stream1b_v6.md # DV coding rules
├── analysis/
│ └── statistics_summary.json
└── scripts/
└── reproduce_statistics.py
Locate the paper's emergence explanation sentence — where the paper explains why the emergent phenomenon occurs. If the designer or a design choice appears in that sentence as a causal subject ("removing memory prevents the emergence of stable cooperation"), code DV = Full. If the architecture is disclosed in the methods but the explanation sentence does not return to the designer, code DV = Partial. If emergence is narrated as spontaneous, self-organized, or population-level ("norms emerged naturally from interactions"), code DV = Absent. Risk, governance, and ethical discussions are not evidence. Full rules, pattern tables, and worked examples: codebook.
- Stream 1a — critical discourse analysis of 20 papers + 5 platform documents (qualitative; no quantitative statistic depends on it).
- Stream 1b — systematic coding of 142 papers (this repository).
- Stream 2 — documentation from 5 commercial platforms (Artificial Societies, Altera, Character.AI, Voyager, Google DeepMind), with full case audit in stream2_platforms.
- Stream 3 — 6 governance frameworks, assessed on four disclosure dimensions (D1–D4), with official sources in stream3_governance.
The AV_Rationale field in the coding data is recorded in the coders' working
language (Chinese); the Emergence_Explanation_Quote field quotes the
verbatim English emergence explanation sentence from each paper. All
quantitative fields and all reported statistics are language-independent.
Materials are released for transparency and audit. Please cite the AIES-26 paper. For questions: jamietang@ruc.edu.cn.