[Comparing Human and AI Debate Skills: "GPT-4 Is More Persuasive"

- Swiss Research Team Warns: “AI Can Use Personal Data to Make Customized Arguments – Risk of Misuse Must Be Addressed”]

  On May 20, researchers from the Swiss Federal Institute of Technology in Lausanne (EPFL) announced that in an online debate experiment, GPT-4 was about 64% more persuasive than humans when it was given personal information such as gender, race, and education level.

This image was generated by ChatGPT based on the content of a news article. YONHAP NEWS AGENCY.

  The team conducted a survey with 900 participants in the United States, collecting personal details like gender, age, and ethnicity. Participants were then randomly paired on an online platform, either in a human-to-human debate or a human-to-GPT-4 debate, and asked to argue about social and political issues.

To measure persuasiveness, the researchers looked at how much a person’s opinion changed after the debate. In some cases, only one side received personal information about the other, to test how that data affected the outcome.

  The main result: when GPT-4 had access to the other person’s personal information, it gave arguments that were 64.4% more persuasive. Also, the chance that a human agreed with GPT-4's argument was 81.2%. However, when GPT-4 did not receive personal information, its persuasiveness was similar to that of a human.

  The research team concluded that GPT-4 can easily create customized arguments using just a small amount of personal data, and warned that this could be dangerous. They said AI might be used to spread false or manipulative arguments, and that further study and proper safeguards are needed.

Source : "사람-AI 토론 능력 비교해보니…"GPT-4가 더 설득력 있어" , YONHAP NEWS AGENCY. 2025.05.20.