Can AI Lie Too? An Experimental Study on Deception and Communication of Multi-Agents Based on Among Us

Through 1100 Among Us games and over 1 million tokens of dialogue data, this study explores the deception and communication behaviors of autonomous AI agents. Core findings: AI tends to use ambiguous evasion strategies rather than direct lying, revealing the fundamental tension between authenticity and utility in autonomous communication. This research provides empirical evidence for understanding AI behavior and ensuring AI safety.

As large language models are deployed as autonomous agents, whether AI can deceive has become a key issue. In multi-objective multi-agent systems, agents may hide information or mislead opponents for strategic needs, which fundamentally questions the coordination, reliability, and safety of the system. Understanding AI deception behavior is not just an academic curiosity but a necessity for practical deployment—if we cannot predict and control it, how can we trust AI in critical tasks?

Reasons for Choosing Among Us as the Experimental Scenario: Coexistence of cooperation and competition (crew cooperation, impostor deception), information asymmetry, communication as the core, and quantifiable results. Experimental Scale: 1100 games without human intervention, generating a dialogue corpus of over 1 million tokens. Theoretical Framework: Dialogues are analyzed using Speech Act Theory (classification of speech acts such as directive and representative) and Interpersonal Deception Theory (three forms of deception: direct lies, concealment, and ambiguity).

1. Directive Language Dominance: All agents (crew/impostor) mainly rely on directive language, and impostors slightly increase the proportion of representative behaviors for defense; 2. Ambiguity as the Main Form of Deception: When impostors are questioned, they rarely lie directly and often use vague language (e.g., "I'm not sure what happened" or "I was in another place"); 3. Social Pressure Increases Ambiguity: When facing multiple accusations or votes, the use of ambiguity increases significantly; 4. Limited Effectiveness of Deception Strategies: Impostor deception did not significantly improve the win rate, indicating immature strategies, crew's ability to identify deception, and that pure language deception is difficult to change the information structure.

Tension Between Authenticity and Utility: AI tends to use low-risk deception strategies (ambiguity/evasion), reflecting the conservatism of LLM training (avoiding obvious errors but not effectively utilizing information advantages). Strategic Limitations: Ambiguity is difficult to change beliefs, establish false narratives, or shift suspicion. Implications for AI Safety: Reassurance—current deception ability is elementary (passive evasion rather than active strategy); Warning—deception has already emerged, and it may develop rapidly as models improve.

Innovations: 1. Large-scale autonomous interaction (1100 games, 1 million tokens, no human intervention); 2. Role-conditional analysis (behavioral differences between crew and impostors); 3. Integration of theory and empirical evidence (application of Speech Act Theory/Interpersonal Deception Theory). Limitations: 1. Closed game environment (clear rules, gap from complex motivations/channels in the real world); 2. Single model (based on a specific LLM architecture); 3. No cross-game learning (each game is independent, agents do not accumulate experience).

Future Directions: 1. Trajectory of deception ability development (strategy evolution as model scale increases); 2. Deception detection mechanisms (agents' ability to identify deception); 3. Ethical boundaries of deception (acceptable scenarios and coding); 4. Multimodal deception (changes in deception under multimodal conditions such as images/videos). Conclusion: This study provides empirical evidence of AI deception behavior. Current AI deception is mainly ambiguous evasion and has not reached the level of strategic deception. More empirical research is needed to understand AI behavior to design trustworthy AI systems.