Have you ever felt subtly swayed by an ad or a piece of writing? That's the power of persuasive language at play. Now, imagine that same power wielded by artificial intelligence. A new research paper explores the fascinating and potentially concerning ability of Large Language Models (LLMs) to generate persuasive text. Researchers delved into this by creating pairs of texts—an original and an LLM-generated paraphrase—designed to be more or less persuasive. These pairs were then judged by human annotators who rated the relative persuasiveness of each. This process resulted in a new dataset, PERSUASIVE-PAIRS, offering a rich resource for understanding how LLMs craft influential language. To go beyond individual judgments, the researchers trained a model to predict persuasiveness scores, creating a benchmark for evaluating different LLMs and their settings. Intriguingly, they found that LLMs, when simply asked to paraphrase without explicit instructions about persuasion, often toned down the persuasiveness of already persuasive text. However, the 'persona' given to an LLM through its system prompt plays a significant role. For example, instructing an LLM to write like a 'tabloid journalist' resulted in significantly more persuasive text compared to a 'scientific journalist.' Similarly, the political leaning assigned to an LLM also affected the persuasiveness of its output. This research highlights the surprising power of LLMs to not only understand but also generate language that influences us. While it opens doors for positive applications, it also raises important ethical questions. As LLMs become more integrated into our lives, understanding how they shape our beliefs and opinions becomes crucial. Future research might explore how different cultures perceive persuasive language and examine the specific linguistic techniques LLMs employ to sway their audience.