On the conversational persuasiveness of large language models: A randomized controlled trial
Authors: F Salvi, MH Ribeiro, R Gallotti
Published: 2024
Publication: arXiv preprint arXiv ..., 2024 - atelierdesfuturs.org
The study demonstrates that GPT-4 is highly persuasive in direct conversations, especially when equipped with personalized sociodemographic information about its opponent, raising concerns about its potential misuse in personalized persuasion contexts.
Methods: Participants engaged in multiple-round debates on a web-based platform under randomized conditions, with comparisons between human-human and human-AI interactions and the impact of personalization.
Key Findings: The persuasiveness of GPT-4 compared to humans, with and without personalization using sociodemographic data.
Limitations: The study is limited to a controlled, harmless debating context, and results may not generalize to real-world persuasion scenarios or diverse populations.
Institution: EPFL, Fondazione Bruno Kessle
Research Area: Conversational Persuasion in LLM
Discipline: Artificial Intelligence
Sample Size: 820 participants
Citations: 118