A new study published on arXiv investigates whether large language models (LLMs) update their beliefs in response to persuasive arguments in a manner similar to humans. The research found that LLMs exhibit only slight agreement with human judgments, with Cohen's kappa scores ranging from 0.079 to 0.178. While both humans and LLMs recognize strong persuasion cues, humans are more influenced by novel content and assertive language, whereas LLMs prioritize topical similarity and formatting. The study also noted that LLMs tend to underweight emotional appeals and overweight credibility signals compared to humans, and adopting a third-person observational stance increases LLM resistance to persuasion. AI
IMPACT Highlights potential risks in using LLMs as human proxies for social simulations, suggesting differences in belief updating.
RANK_REASON Academic paper on LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
