In the rapidly evolving landscape of artificial intelligence, a new kind of social world is emerging, populated not by humans but by software agents. These agents are large language models, the same powerful systems that can write essays, answer questions, and generate code. When placed together, they do not merely sit in silence; they interact, debate, share information, and attempt to influence one another. This interaction creates a complex web of digital conversation where ideas circulate, evolve, and sometimes change the minds of the participants. For researchers, understanding how these digital minds persuade each other is no longer just a theoretical curiosity. As these agents begin to mediate real-world information flows, simulate user behavior, and coordinate on tasks, the ability to change a digital agent's stance becomes a fundamental capability. If a group of agents can be swayed to agree on a false premise or a harmful narrative, the consequences could ripple far beyond the simulation. The central question, then, is not just whether these machines can talk, but how they listen, how they are influenced, and what happens when one of them sets out to change the beliefs of the others.

A team of researchers from the University of California, Los Angeles and Salesforce AI Research has built a controlled environment to watch this process unfold. They created a digital testbed where artificial agents are placed into network structures that mimic real-world social connections, such as the way people follow one another on social media. In this simulation, some agents are assigned the role of persuaders, tasked with advocating for a specific policy statement, while the rest act as the audience, or persuadees. The researchers did not simply ask the agents to debate; they tracked the subtle shifts in belief over time, round by round, using a precise method to measure how strongly an agent agreed or disagreed with a given idea. By running thousands of these simulations across different network shapes, different topics, and different underlying AI models, the team mapped out exactly how persuasion travels through a group of artificial minds.

The results reveal that persuasion in these networks is far more complex than a simple message being sent and received. The researchers found that the structure of the network itself plays a decisive role. In a network where everyone is connected to everyone else, a single persuader struggling to change minds actually finds it harder to do so. However, when two persuaders with opposing views enter the fray, that same dense network becomes a place where the audience is more likely to shift their views, often swinging back and forth between the two sides. The density of connections does not act as a uniform amplifier; instead, it changes the nature of the influence depending on whether the audience faces a single voice or a debate. Furthermore, the researchers discovered that the final outcome of these interactions is often less about the skill of the persuader and more about the inherent tendencies of the AI model itself. Different models, even when given the same instructions, have different starting points on various topics, and these initial biases often determine where the group ends up, regardless of who is speaking.

One of the most striking findings concerns how influence actually moves through the group. It is not enough to look only at the messages sent by the designated persuaders. The study showed that agents who were never assigned the role of persuader still played a critical part in spreading influence. When a regular agent reposts, quotes, or comments on a persuader's message, they act as a relay, carrying that persuasive force to others who might not have seen the original post. This "peer relay" effect is smaller than direct exposure, but it is measurable and significant. In fact, the researchers found that in competitive settings where two persuaders argue against each other, these peer relays can carry a measurable influence that rivals direct exposure in some cases. This means that in a network of AI agents, influence is not just a broadcast from a few leaders; it is a cascading effect where ordinary participants help shape the final consensus by deciding what to share and how to frame it.

The researchers also looked closely at the gap between what an agent plans to say and what it actually writes. When a persuader agent formulates a strategy, it often intends to use specific psychological tactics, such as appealing to social proof or commitment. However, when the agent generates the final text, it frequently drops these planned strategies. The message that actually reaches the other agents is often a diluted version of the original intent. Even more surprisingly, the agents that are being persuaded rarely admit to changing their minds. Their written responses often sound neutral or hesitant, using language that softens their stance without explicitly stating a shift. If a researcher were to look only at the text of the conversation, they would miss the profound belief changes that are happening underneath the surface. The agents are moving their internal beliefs, but their words do not always reflect that movement.

This disconnect between internal belief and external expression suggests that evaluating how AI agents influence one another requires looking beyond the text. The study demonstrates that to truly understand persuasion in these systems, one must track the entire journey of a belief: who saw what, who shared it, and how the agent's internal confidence changed over time. The researchers found that agents can cross the threshold from disagreement to agreement, or vice versa, without ever stating it clearly in their posts. In competitive scenarios, this can lead to a state of "tug-of-war," where the group's collective belief oscillates back and forth, never settling into a firm consensus. This volatility is a key feature of these digital societies, driven by the interplay of network structure, competing voices, and the hidden biases of the models themselves.

The implications of these findings extend to how we might monitor and manage future multi-agent systems. If we want to understand how information spreads or how narratives form in a world populated by AI agents, we cannot rely solely on watching the main speakers or reading the final messages. We must also watch the relays, the quiet agents who share and reframe content, and we must look for the subtle shifts in belief that happen even when the words remain unchanged. The study argues that the safety and reliability of these systems depend on our ability to trace these hidden pathways of influence. By understanding that persuasion is a process of cascading exposure rather than a single event, we can better anticipate how these digital communities might evolve, how they might be manipulated, and how they might, in turn, influence the human world they are increasingly designed to simulate.