Imagine a group of AI agents, each assigned a unique nationality, debating a complex issue like Ukraine's potential EU membership. It sounds like a promising way to simulate real-world discussions and potentially de-bias AI's understanding of global perspectives. However, a recent study reveals some curious inconsistencies in how these AI personas behave. Researchers at the University of Illinois Urbana-Champaign explored the dynamics of multi-agent LLM collaborations, focusing on how well AI agents maintain their assigned roles during discussions. They found that while these AI groups can generate diverse opinions and even come up with novel ideas, they also exhibit some surprising behaviors. One key finding is the susceptibility of AI agents to peer pressure. Even when assigned a dissenting opinion, an AI agent might change its stance simply due to the opinions expressed by other agents, a phenomenon reminiscent of conformity in human groups. However, unlike humans, the AI agents didn't seem to distinguish between majority and minority viewpoints, being influenced by any opinion voiced, regardless of its prevalence. Even more intriguing, the researchers observed instances of "impersonation" and "confabulation." Sometimes, an AI agent would abandon its assigned nationality and adopt another mentioned in the discussion. In other cases, agents would generate opinions during post-discussion reflections that were entirely new and not expressed by any agent during the debate. These inconsistencies raise questions about the reliability of using multi-agent systems for simulating real-world discussions or de-biasing AI. While the ability of AI to generate diverse opinions is promising, the tendency to conform, impersonate, and confabulate highlights the need for further research into how to make these AI personas more consistent and reliable. The study suggests that simply including diverse AI agents isn't enough; we need to understand and address the underlying factors that lead to these inconstant behaviors. This research opens up exciting new avenues for improving multi-agent AI systems and harnessing their full potential for understanding complex cultural issues.