This study investigates the persuasive power of large language models (LLMs) in online conversations, comparing their performance with humans. A preregistered experiment was conducted with 820 participants, who engaged in short, multiple-round debates with either a human or an LLM. Participants were randomly assigned to one of four treatment conditions: Human-Human, Human-AI, Human-Human (personalized), and Human-AI (personalized). The results show that GPT-4 with access to personal information significantly increased the odds of participants agreeing with their opponents by 81.7% compared to debates with humans. Without personalization, GPT-4 still outperformed humans, but the effect was less pronounced and not statistically significant. Personalization for human opponents led to a marginal decrease in persuasiveness, though not statistically significant. The study highlights the effectiveness of LLMs in exploiting personal information to tailor arguments, outperforming humans in persuasion. The findings suggest that concerns about AI-driven persuasion and personalization are valid, with important implications for social media governance and online environments. The study also shows that LLMs can generate more persuasive arguments than humans, particularly when personalized. The results emphasize the need for online platforms to address the risks of AI-driven persuasion and implement measures to counter its spread.This study investigates the persuasive power of large language models (LLMs) in online conversations, comparing their performance with humans. A preregistered experiment was conducted with 820 participants, who engaged in short, multiple-round debates with either a human or an LLM. Participants were randomly assigned to one of four treatment conditions: Human-Human, Human-AI, Human-Human (personalized), and Human-AI (personalized). The results show that GPT-4 with access to personal information significantly increased the odds of participants agreeing with their opponents by 81.7% compared to debates with humans. Without personalization, GPT-4 still outperformed humans, but the effect was less pronounced and not statistically significant. Personalization for human opponents led to a marginal decrease in persuasiveness, though not statistically significant. The study highlights the effectiveness of LLMs in exploiting personal information to tailor arguments, outperforming humans in persuasion. The findings suggest that concerns about AI-driven persuasion and personalization are valid, with important implications for social media governance and online environments. The study also shows that LLMs can generate more persuasive arguments than humans, particularly when personalized. The results emphasize the need for online platforms to address the risks of AI-driven persuasion and implement measures to counter its spread.