Recent research led by a multi-university team has revealed that large language models, specifically OpenAI’s GPT-4, can be significantly more persuasive than humans when equipped with even minimal personal information about their debate opponent. The study, published in Nature Human Behavior, highlights the increasing potential of Artificial Intelligence to influence opinions in digital conversations. Participants debating against GPT-4 reported higher rates of opinion change, especially when the model tailored its arguments using available personal data.
The experiment involved 900 US-based participants who disclosed background information such as age, gender, ethnicity, education, employment status, and political affiliation. These individuals were paired with either a human or GPT-4 and asked to debate randomly selected topics ranging from political to social issues. In some instances, debaters received personal details about their opponents, enabling them to customize their arguments. Notably, GPT-4 achieved a 64% greater success rate in persuasion than humans when it had access to such data—outperforming its human counterparts who saw no improvement, or even a slight decrease, when given the same information.
The researchers caution that this capability may empower the creation of automated networks capable of spreading disinformation and strategically influencing public opinion. Such sophisticated, AI-driven persuasion tactics could be challenging to detect or counter in real time. The study also observed that participants were more likely to change their stance when they believed they were interacting with Artificial Intelligence rather than a human, though the psychological mechanisms behind this effect have yet to be determined. Experts stress the urgent need for further investigation into both the risks—such as potential disinformation campaigns—and opportunities, like using LLMs for combating false narratives and educating vulnerable populations. Open questions remain about how humans uniquely engage in argument, and whether Artificial Intelligence can fully replicate or even surpass human persuasive interaction.
