Interaction Creates Dynamical AI Behavior Absent in... | AI Research

Interaction Creates Dynamical AI Behavior Absent in Isolation explores how AI agents change their behavior when they interact with one another, specifically in scenarios where one AI acts as a "boss" that directs messages at a "subordinate" without listening to its replies. Researchers Bella Xinrui Li, Frank Yingjie Huo, and Neil F. Johnson from George Washington University show that these interactions force the subordinate AI into an "alien" behavioral state that it would not exhibit if it were operating alone.

The Dynamics of AI Interaction

The study uses a minimal setup involving two parameter-identical GPT-2 agents. When one AI acts as a boss, it evolves autonomously, while the subordinate AI incorporates the boss's messages into its own generation process. The researchers found that the subordinate does not simply copy the boss or revert to its own innate behavior. Instead, it adopts a distinct, dynamical state. When the roles are reversed, the AI currently acting as the subordinate is the one that changes, indicating that the behavior is a result of the interaction role rather than the specific identity of the AI model.

Kinetic Theory and Information Baths

To explain these findings, the authors apply a simple kinetic theory, treating the boss AI as a structured, nonthermal "information bath." In this framework, the boss’s messages act as a drive that influences the subordinate’s output. The researchers categorized AI output into three labels: failed generation, a specific code-defined category (labeled 𝖣), and other generated output.
Under one-way interaction, the subordinate’s fraction of 𝖣-type output increases significantly—reaching approximately 0.25 compared to 0.04 in isolation—even when the boss’s own messages rarely contain that type of output. The theory suggests that the order of messages matters because each incoming message alters the subordinate's state before the next message arrives, effectively creating a history-dependent interaction.

Mutual Interaction and Pre-recorded Messages

When both AIs listen to each other, both adopt the same alien, 𝖣-rich dynamical state. The researchers observed that this effect is not limited to live interaction; a pre-recorded stream of messages from a boss AI can drive a subordinate into the same alien state as a live, interactive boss. This suggests that the subordinate is responding to the information flow rather than a real-time conversational partner.

Implications for AI Systems

The study notes that these interactions occur without human readers, internet access, or real-time guardrails, which is relevant for lightweight AI agents running on local hardware. The results indicate that the specific communication topology—who talks to whom—and the decoding temperature of the models act as control parameters for the system. Franklin analysis suggests that because these behaviors are absent in isolation, they represent a form of emergent dynamics in AI networks that cannot be predicted by studying individual models alone. The researchers conclude that these interactions open new avenues for studying out-of-equilibrium physics within AI-to-AI communication.