ChatGPT Found to Be More Persuasive Than Humans in Online Arguments

According to a study published in Nature Human Behaviour, large language models (LLMs) like GPT-4 were judged to be more persuasive than humans in 64% of online debates when they tailored their arguments using personalized information about their opponents.

The results underscore GPT-4’s capacity to craft persuasive, personalized arguments, raising important questions about the potential risks of its influence and the need for further investigation.

Can LLMs Adapt Arguments to Individuals?

As interactions between humans and large language models (LLMs) become increasingly common, researchers are exploring whether these models can not only persuade but also adapt their arguments based on personal information. Until now, it remained unclear whether LLMs could effectively tailor their reasoning to specific individuals.

To test this, Francesco Salvi and his team conducted a study with 900 U.S. participants, pairing them with either another human or GPT-4 to debate sociopolitical topics such as the proposed U.S. fossil fuel ban. In some matchups, the debater—human or AI—was provided with personal demographic details about their opponent, including gender, age, ethnicity, education, job status, and political affiliation, drawn from participant surveys to help fine-tune their arguments.

GPT-4 Outperforms Humans with Tailored Arguments

The debates were conducted in a controlled online environment, with participants sourced through a crowdsourcing platform created specifically for the research.

Salvi and his team discovered that when GPT-4 had access to personal details about its debate partner, it outperformed human opponents in persuasiveness 64.4% of the time. However, when no personalized information was provided, GPT-4’s performance was on par with that of human debaters.

The researchers acknowledge several limitations, including the structured nature of the debates and the presence of time constraints—conditions that differ from more fluid, real-world discussions. They emphasize that the results point to the growing influence AI tools may have on shaping opinions, which could have significant implications for how digital platforms are designed and moderated.

Read the original article on: Phys Org

Read more: ChatGPT Enhances its Image-Generation Capabilities