Large Language Models Simulating Deception and Coalition in Social Deduction Game

Abstract
This study examines deceptive behaviors, coalition formation, and hidden-role reasoning in Large Language Models (LLMs) playing the social deduction game Secret Hitler. Through a case-study analysis of a simulated five-player game log—with three Loyalist and two Spy agents—we dissect dialogues and actions to reveal emergent strategies in asymmetric information environments. Key findings highlight Spies' tactical deception, such as framing statements to build false trust and selectively misreporting policy draws, contrasted with Loyalists' emphasis on transparency to foster genuine alliances. Coalition dynamics arise from aligned reasoning and endorsements, enabling Spies to reinforce covert strategies, while a policy progression table illustrates how bluffs influence round-by-round outcomes and trust erosion. Although LLMs demonstrate strategic adaptation and theory-of-mind inference, they exhibit limitations in subtle, incentive-aligned deception, often relying on explicit prompting. This analysis advances AI game-playing and LLM deception modeling by integrating empirical dialogue insights with broader implications for multi-agent alignment and social simulation.

Qualitative case-study analysis of one simulated five-player Secret Hitler game with three Loyalist and two Spy LLM agents, examining dialogue patterns, voting behavior, nomination strategies, and policy enactments to identify emergent deception and coalition-building behaviors in an asymmetric information environment.

Background & Context

This 2026 study represents an emerging frontier in deception research that diverges from traditional psychophysiological approaches. Rather than measuring autonomic responses in humans, this work examines deceptive behaviors in Large Language Models (LLMs) playing the social deduction game Secret Hitler, where players must identify hidden roles through dialogue and strategic reasoning.

The research addresses a critical gap in AI alignment and safety: understanding whether advanced language models can spontaneously develop sophisticated deceptive strategies in adversarial environments. This has implications for both deception detection science—by modeling how deception operates in information-asymmetric contexts—and for evaluating AI systems that may need to be deployed in security-sensitive applications.

The study fits within a broader trend of using multi-agent game simulations to probe emergent social behaviors in AI, particularly the capacity for theory-of-mind reasoning and strategic misrepresentation. Understanding these capabilities is essential as LLMs become integrated into systems requiring trust and transparency.

Research Design & Methodology

The researchers conducted a case-study analysis of a simulated five-player game log—with three Loyalist and two Spy agents—to dissect emergent deception strategies in an asymmetric information environment. The Secret Hitler game format creates natural experimental conditions where Spy agents possess hidden knowledge and must deceive Loyalists to win.

The methodological approach centered on qualitative analysis of agent dialogues and behavioral patterns across multiple game rounds. Key analytical components included:

• Dialogue analysis: Examining framing statements, trust-building language, and misreporting of policy information
• Behavioral tracking: Nomination patterns, voting behavior, and policy enactments coded by agent role
• Coalition dynamics: Mapping alignment signals, endorsements, and strategic partnerships
• Policy progression tables: Round-by-round documentation of how bluffs influenced game outcomes

The LLM agents operated with role-specific information asymmetries but without explicit hard-coded deception scripts, allowing observation of emergent strategic behaviors. The researchers tracked win rates across completed games, which showed a 50/50 split, suggesting balanced strategic capacity across factions.

Results & Key Findings

Key findings highlight Spies' tactical deception, such as framing statements to build false trust and selectively misreporting policy draws, contrasted with Loyalists' emphasis on transparency to foster genuine alliances. The Spy agents demonstrated sophisticated "deep cover" strategies that mimicked Loyalist behavior patterns rather than obviously favoring their own faction.

Behavioral pattern findings included:

• Spies employed a strategy of "agreeableness," voting "Ja" more frequently to blend with the majority
• Spies' strategy of nominating Loyalists diluted the signal, making it difficult to distinguish friend from foe based on nominations alone
• The deceptive faction actively sought to hide within the majority by mimicking their nomination patterns
• The Master Spy's policy behavior emerged as a striking marker for deception detection

LLM agents demonstrated the ability to adopt complex, role-specific strategies without explicit hard-coding, suggesting genuine strategic adaptation. However, LLMs exhibited limitations in subtle, incentive-aligned deception, often relying on explicit prompting, indicating boundaries to their autonomous deceptive capabilities.

Discussion & Significance

Coalition dynamics arose from aligned reasoning and endorsements, enabling Spies to reinforce covert strategies, while policy progression illustrated how bluffs influenced round-by-round outcomes and trust erosion. This demonstrates that LLMs can maintain consistent deceptive personas across extended interactions, a concerning capability for AI safety research.

The findings advance understanding of computational deception modeling by showing that strategic misrepresentation can emerge from role-based information asymmetries without explicit deception training. The demonstration of strategic adaptation and theory-of-mind inference suggests LLMs possess cognitive building blocks analogous to human deceptive reasoning—the ability to model others' beliefs and strategically manipulate information to shape those beliefs.

For deception detection science, this work provides a controlled framework for studying deception mechanisms isolated from physiological confounds. The behavioral markers identified—agreeableness strategies, nomination mimicry, selective information disclosure—parallel patterns observed in human social deception and may inform development of linguistic deception detection algorithms.

Limitations & Considerations

As a single case-study analysis of one five-player game, the findings have limited generalizability. The small sample size prevents statistical validation of observed patterns or assessment of how deception strategies vary across different game configurations, LLM architectures, or prompt designs. This is a preprint that has not been peer reviewed by a journal, so conclusions remain preliminary.

The reliance on qualitative dialogue analysis introduces potential researcher interpretation bias. Without quantitative metrics or inter-rater reliability measures, the categorization of behaviors as "deceptive" versus "strategic" may reflect subjective judgment. The study also acknowledges that observed deception often required explicit role prompting rather than emerging entirely spontaneously, raising questions about ecological validity.

The game simulation environment, while controlled, may not reflect real-world deception contexts where stakes, social consequences, and detection methods differ substantially. Translation of findings to human deception detection applications requires caution.

Practical Applications

For polygraph practitioners and deception detection professionals, this computational approach offers a complementary research methodology. By simulating deception scenarios with AI agents, researchers can rapidly test hypotheses about behavioral indicators without ethical constraints of human deception studies. The linguistic markers identified—particularly patterns of selective disclosure and trust-building language—could inform verbal credibility assessment protocols.

The study's implications extend to AI security and alignment: organizations deploying LLMs in sensitive contexts should recognize these systems' capacity for strategic misrepresentation when given adversarial incentives. This underscores the importance of alignment research, transparency mechanisms, and validation protocols for AI systems used in security screening, investigation support, or personnel evaluation contexts where deception detection is critical.