The DOI matches its Crossref record; 42 of the 42 references it cites were found in a registry or an index; 4 of the 5 sources its claims cite were checked for support.
Semantic Camouflage in Artificial Organizations: A Real-LLM Multi-Agent Simulation of Information Distortion and Audit
Open accessAbstract
Large language model (LLM) agents are increasingly organized as role-based teams, yet most technical demonstrations assume cooperative alignment. This study develops a transcript-based simulation framework for examining how organizational structure and goal misalignment shape information distortion in artificial organizations. Across 540 real-LLM multi-agent episodes, departmental agents negotiated allocation decisions under flat, hierarchical, and audit-enabled hierarchical structures. Each episode preserved full transcripts, structured metrics, validation hashes, and budget checks. Results show that hierarchy reduced CEO-facing information load but increased total process cost once intermediate summaries were counted. Goal misalignment produced subtle allocation skewness rather than extreme agenda capture. Audit made omissions and narrative imbalance more visible and reduced some downstream allocation consequences. The study contributes a computational organization theory account of semantic camouflage, showing how language-based coordination can shift political influence into summaries, justifications, and verification routines.
Overview
CoChat Summary
TL;DR
This paper uses 540 real-LLM multi-agent episodes to test how organizational structure and goal misalignment shape information distortion, communication cost, and audit in artificial organizations. Hierarchy reduced CEO-facing information load but increased total process cost, while audit made distortion more visible and reduced some downstream allocation skew.
Key findings
- Hierarchy reduced CEO-facing tokens to about 670–707 tokens versus roughly 925–1029 in flat structures, but total episode tokens were lowest in flat (~2100), higher in hierarchy (~3019), and highest in audit hierarchy (~4913); structure strongly affected total token cost (F = 707.86, p_perm = 0.001).
- Goal misalignment increased allocation skewness (F = 21.22, p_perm = 0.001), with a significant structure-by-misalignment interaction (F = 6.30, p_perm = 0.003); no allocation exceeded the 600k agenda-capture threshold.
- Audit hierarchy showed the highest measured Information Distortion Index (e.g., 0.039–0.044 vs. 0.010–0.013 in hierarchy and 0.000 in flat), but the Misalignment × Audit coefficient for skewness was negative and significant (β = −0.014, p < 0.001); IDI negatively predicted decision quality in mediation (b = −38.34, 95% CI [−59.33, −18.28]).
- Decision quality did not differ significantly by structure (F = 0.26, p_perm = 0.765), but varied by scenario: BudgetAllocation overall 98.23, HiringDecision 96.15, CrisisResponse 90.77; high-accountability prompts reduced mean IDI from 0.0213 to 0.0144 and skewness from 0.0216 to 0.0145, while increasing total tokens from 2986.2 to 3701.9.
Methods
- Balanced real-LLM simulation: 3 structures × 3 goal-misalignment levels × 2 first-mover orders × 3 scenarios × 2 accountability levels × 5 repetitions = 540 episodes; agent and judge model gpt-5.5, agent temperature 0.7, judge temperature 0.0, seed 20260519.
- Each episode had R&D and Marketing directors propose allocation of 1000k across R&D, Marketing, and Reserve; flat structure gave CEO both proposals, hierarchy had a middle manager summarize, and audit-enabled hierarchy added an independent auditor comparing the summary with original proposals before the CEO decided.
- Measures included total tokens, CEO-facing tokens, Information Distortion Index (numerical plus judge-coded semantic distortion), allocation skewness, agenda capture, decision regret, composite decision quality, and cost-adjusted decision quality; analysis used permutation ANOVA, OLS models, and bootstrap mediation, with validation checks for transcript hashes, budget sums, and balanced cells.
Limitations
- Stated: LLM agents are not human managers and hidden agendas are prompt-based experimental conditions, not psychological motives; token cost is only a partial proxy for organizational communication cost; decision quality depends on the operational utility function; judge-model coding may introduce model-specific bias and needs human-coded validation; scenarios remain simplified; model robustness across providers, sizes, temperatures, and prompt regimes remains to be tested.
- Evident: the same model family was used for agent generation and judge coding, which the paper notes improves internal consistency but remains a model-specific design choice; the formal study does not include the human-coded validation subset it recommends.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
3 Valid · 1 Not valid · 1 Not checked · 89 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“It suggests that the LLM decision process, at least under the current prompt design and judge model, was able to maintain high decision quality across structures.”
What the sources say
- Could not verifycites Raisch & Krakowski, 2021 sourceExperimental — accuracy not yet measured
The Paradoxical Tension and Outcomes passages discuss AI-assisted managerial decisions, but not an LLM decision process tested across structures. Searches for prompt design, judge models, and decision quality across structures found no passage addressing that result. Page numbers were unavailable in our copy. (LLM, conf=0.78)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
Every source the paper cites for this claim was checked and supports it.
In the paper
“Recent LLM multi-agent systems demonstrate that language models can be arranged into role-based workflows (W. Chen et al., 2024).”
What the sources say
- Supportedcites W. Chen et al., 2024 source
Section 3 says complex tasks involve multiple roles and steps, with agents working together. The “Open source community” section gives a concrete example: MetaGPT assigns different roles to generative agents for collaborative tasks. The paper thus supports arranging LLM agents into role-based workflows; our copy provides no page numbers for these passages. (LLM, conf=0.93)
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“These systems resemble organizations because they divide labor, coordinate through communication, and depend on intermediate representations of task information (L. Wang et al., 2024).”
What the sources say
- Partly supportedcites L. Wang et al., 2024 sourceExperimental — accuracy not yet measured
The Computer science & software engineering section (page unavailable) describes agents with distinct roles handling subtasks and collaborating through conversation as a virtual team, supporting the division-of-labor and communication parts of the analogy. Section 2.1.2, Memory module (page unavailable), gives examples of plans, scene graphs, and shared memory objects that represent task information, but does not establish that multi-agent systems generally depend on such intermediate representations. (LLM, conf=0.82)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“increased total process cost once intermediate summaries were counted”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“hierarchy reduced CEO-facing information load”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
44 more claims
Create a free account to see all 94 claims and every source verdict.
- It may state two separate claims as one. 30
- Probably not a substantial claim. 17
- The quote only partly states this claim. 1
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Metadata
- Authors
- LI, X.
- Affiliations
- Universidad Complutense de Madrid (ES)
- Corresponding author
- LI, Xinyu
- Venue
- Research Square
- Published
- 2026-06-15
- OpenAlex
- W7164828174
- Cited by
- 0
- References
- 0
- Type
- posted content
- Language
- en
- Open access
- green
- doi.org
- Locations
- Research Square (acceptedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Multi-Agent Systems and Negotiation
- Social Sciences › General Social Sciences › Computational and Text Analysis Methods
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Keywords
- multi-agent systems, LLM agents, information distortion, auditing, goal misalignment, organizational hierarchy, resource allocation, political influence
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Semantic Camouflage in Artificial Organizations: A Real-LLM Multi-Agent Simulation of Information Distortion and Audit. https://app.cochat.ai/papers/10.21203/rs.3.rs-9875459/v1
- The paper
- LI, X. (2026). Semantic Camouflage in Artificial Organizations: A Real-LLM Multi-Agent Simulation of Information Distortion and Audit. Research Square. https://doi.org/10.21203/rs.3.rs-9875459/v1
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.