Imagine a game of telephone, but instead of whispering secrets, we have powerful AI language models rewriting text. What happens to the information as it passes from one AI to another? Researchers explored this intriguing question in a recent study. They set up a virtual game of telephone, using different LLMs like ChatGPT and Llama. One AI would receive a text and be instructed to rephrase it, take inspiration from it, or continue it. The output was then passed to the next AI in the chain, and the process repeated. The results revealed fascinating dynamics. Some AI models showed a tendency to strip away negativity, making the text progressively more positive. Others altered the complexity, sometimes simplifying and sometimes making the text more intricate. The format of instructions also played a key role. Open-ended tasks like 'continue the story' resulted in greater transformations compared to stricter tasks like 'rephrase.' The research also highlighted how these cumulative changes can lead to "attractor states." This means that, despite starting with diverse texts, the AI telephone game often converges towards similar themes or styles. Imagine several groups playing telephone with the same starting phrase. After several rounds, each group's final phrase will be different but likely share some similarities. These similarities represent the attractor states. This research has significant implications for how we understand and use LLMs. As AI-generated content becomes more prevalent, understanding how information mutates as it passes through multiple AIs is critical. The research suggests that single-turn evaluations of LLMs might not be enough. We need to consider how they interact and influence each other in a chain, particularly in applications like simulating social interactions or building AI-driven narratives. This is just the beginning of understanding the complex cultural evolution within AI systems, opening exciting new avenues for research and shaping how we use these tools responsibly.