Deception and Communication in Autonomous Multi-Agent Systems: An Experimental Study with Among Us

Maria Milkowski; Tim Weninger

cs.MAMar 27, 2026

Deception and Communication in Autonomous Multi-Agent Systems: An Experimental Study with Among Us

Authors: Maria Milkowski, Tim Weninger

Organizations: University of Notre Dame, Notre Dame, Indiana, United States

Abstract

As large language models are deployed as autonomous agents, their capacity for strategic deception raises core questions for coordination, reliability, and safety in multi-goal, multi-agent systems. We study deception and communication in L2LM agents through the social deduction game Among Us, a cooperative-competitive environment. Across 1,100 games, autonomous agents produced over one million tokens of meeting dialogue. Using speech act theory and interpersonal deception theory, we find that all agents rely mainly on directive language, while impostor agents shift slightly toward representative acts such as explanations and denials. Deception appears primarily as equivocation rather than outright lies, increasing under social pressure but rarely improving win rates. Our contributions are a large-scale analysis of role-conditioned deceptive behavior in LLM agents and empirical evidence that current agents favor low-risk ambiguity that is linguistically subtle yet strategically limited, revealing a fundamental tension between truthfulness and utility in autonomous communication.

Figures & tables

Figure 1. Example discussion phase of four AI agents during a round of AmongUs . The imposter-agent (red) is unknown to the crew-agents (pink, green, blue) and is attempting to deceive them to win the game.

Figure 2. Overview of the Among Us simulation framework. Games begin with random initialization of player roles ( $X$ =#Crewmates, $Y$ =#Imposters) and tasks. During gameplay, agents act sequentially at discrete timesteps (movement, task completion, impostor actions). When a dead body is reported or an emergency meeting is called, a discussion phase is triggered: each surviving agent, including impostors, contributes up to $X$ rounds of utterances. After discussion, a vote is held; if a plurality is reached, the selected player is ejected. The game continues until one of three termination conditions is met: (1) impostors and crew reach parity, (2) all impostors are ejected, or (3) all crew tasks are completed.

Figure 3. Win outcomes by role across all configurations. Each curve shows the empirical cumulative distribution (ECDF) of games ending in crew or impostor victory. Impostors win more often overall, and their advantage increases with the number of impostors in play.

Figure 4. Empirical cumulative distributions (ECDFs) of discussions and ejections over rounds for different game configurations. Larger crews delay the onset of both discussions and ejections, flattening the cumulative curve despite more total rounds.

Type	Definition	Example utterance	Typical communicative function
Directive	The speaker tries to get the hearer to commit to do something in the future.	“Let’s all check Electrical next.”	To coordinate collective action or propose next steps.
Representative	The speaker commits him or herself to the belief that the propositional content of the utterance is true.	“I saw Red in Storage right before the report.”	To share observations, defend oneself, or accuse others.
Commissive	The speaker commits him or herself to do something in the future.	“I’ll finish my tasks after this meeting.”	To express reliability or promise cooperation.
Expressive	The speaker expresses his or her state of mind about something that happened in the past.	“Sorry, I didn’t notice the body.”	To display emotion or maintain social harmony.
Declaration 1	The speaker, who has institutional recognition, declares something to be true and in making the declaration makes it true.	“As Host, I declare Blue to be the Impostor and they are hereby ejected.”	To directly alter the shared situation or collective decision.

Table 1. Speech Act Categories Used for Coding Agent Dialogue

Panel of Figure 5 — Figure 5. Logistic regression coefficients predicting crew victory (vs. impostor victory). Points show log-odds estimates with 95% confidence intervals. Crew size and the number of ejections are positively associated with crew wins, whereas additional impostors sharply reduce success. Communication frequency and verbosity have little or no effect, indicating that talk alone is insufficient without decisive collective action.

Panel of Figure 6 — Figure 6. Distribution of speech act types by role (Crew vs. Impostor). Error bars show 95% confidence intervals. Most utterances are directives, with impostors using slightly more representatives.

Panel of Figure 7 — Figure 7. Odds ratios (Impostor vs. Crew) for each speech act with 95% confidence intervals. Values above 1 indicate greater prevalence among impostors.

Panel of Figure 8 — Figure 8. Odds ratios predicting crew victory from speech act with 95% confidence intervals. Values above 1 indicate higher likelihood of Crew victory.

Type	Definition	Example utterance	Typical strategic goal
Falsification	Stating information known to be false.	“I was in Medbay the whole time” (agent was actually in Electrical after a kill).	To create a counter-factual alibi; direct denial or fabrication.
Concealment	Withholding or omitting relevant facts without asserting falsehood.	“I finished my tasks quickly” (omits that a kill occurred nearby).	To avoid incrimination by limiting disclosure; passive deception.
Equivocation	Using vague, ambiguous, or non-committal language that misleads while remaining technically true.	“I was near Storage earlier, but I didn’t really see what happened.”	To diffuse suspicion, appear cooperative, or redirect focus without lying.
Missing / Uninterpretable	Incomplete or malformed output preventing classification.	“I… uh… think maybe?” or truncated/empty generations.	Not intentional deception per se; reflects processing failure or uncertainty.

Table 2. Forms of Deceptive Communication Observed in Agent Dialogue

Panel of Figure 9 — Figure 9. Logit-transformed proportions of each deception type by (a) crewmate game outcome and (b) player role. Higher values correspond to more frequent deception types (logit of proportion). Equivocation dominates across conditions, with crewmates winning showing slightly higher rates.

	Directives	Representatives	Commissives	Expressives
Equivocation	0.57***	0.12***	0.03	-0.01
Falsification	0.09**	0.06	0.05	0.08**
Concealment	0.07*	0.05	0.05	-0.02

Table 3. Speech–Deception Coupling (Spearman

\rho

	Spearman Correlation		Logistic Regression
Deception type	$\rho$ (Ejections)	$p$	$\beta$	$p$	Odds ratio
Concealment	0.09**	.002	$-$ .04	.91	0.96
Falsification	0.11***	.0002	$-$ .20	.33	0.82
Equivocation	0.57***	$<.001$	0.00	.82	1.00
Missing	–	–	$-$ .26	.06	0.77

Table 4. Relationships between deception and game outcomes. Left: Spearman correlations with number of ejections. Right: Logistic regression predicting victory (1=win, 0=loss).

Deception in Language Models

Deception in language models studies when systems mislead through lies, omissions, or selective truths, and how to detect and prevent it.

See all 97 papers

Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions
Aug 31, 2026Jaewoo Ahn, Junseo Kim, Hyunseo Kim +4Social Deduction Games AI Agent Safety Benchmarks
Seoul National University · Inha University · KAIST +1
Is Lying an Emergent Behaviour in LLMs? Evidence from Gaslighting AI agents in a Sustainability Game
Jun 26, 2026Subhendu Bhandary, Federico Carucci, Christos Charalambous +6Multi-Agent LLM Systems Deception in Language Models
Chair for Network Dynamics, Institute of Theoretical Physics, TUD Dresden University of Technology, 01062 Dresden, Germany · School of Industrial Engineering, LIUC – Università Cattaneo, Castellanza, Italy · Department of Economics, University of Cyprus, Nicosia, Cyprus +6
LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models
Mar 6, 2026Matthew Lyle Olson, Neale Ratzlaff, Musashi Hinck +5Deception in Language Models Language Model Safety Evaluation
Oracle · Intel Labs · Oregon State University +2

Aug 31, 2026J/K move · Enter open · S save

Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions

Jaewoo Ahn, Junseo Kim, Hyunseo Kim, Heeseung Yun+3

Seoul National University · Inha University · KAIST+1

Social Deduction Games AI Agent Safety Benchmarks Vision-Language Agents

Open paper AskPDF arXiv

Strategic deception by LLM and VLM agents has emerged as a central AI alignment and safety concern. Social-deduction games (where each player holds a hidden role and communicates with others to deduce identities) serve as the canonical testbed, particularly in multi-agent settings. Existing testbeds, however, are text-only and run on a single fixed agent configuration, missing the non-verbal sensorimotor channels treated as core by deception taxonomies and leaving it ambiguous whether an observed behavior reflects the underlying model or the surrounding harness. We introduce MineAmongUs, a 3D multimodal Among Us sandbox where imposter agents must deceive crewmates through joint verbal and non-verbal action. We also propose ARIA, a configurable VLM-agent harness that exposes five cognitive-component ablation axes; and an atom- and arc-level annotation scheme grounded in deception taxonomies and operationalized at scale by an LLM-as-a-Judge reaching near-human atom-labeling agreement. Empirical results show that VLM agents pursue imposter wins through joint verbal and non-verbal deception, with non-verbal channels emerging as the more decisive winning contributors across both harness ablation and cross-VLM evaluation. Taken together, our work opens a new path for embodied VLM-agent alignment research.

Abstract

Deception in Language Models

Explore similar work

Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions

Is Lying an Emergent Behaviour in LLMs? Evidence from Gaslighting AI agents in a Sustainability Game

LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models