Mind What You Ask For: Emotional and Rational Faces of Persuasion by Large Language Models

TL;DR

The paper analyzes how 12 large language models respond to emotional versus rational persuasion prompts, using LIWC-22-derived psycholinguistic metrics and a structured annotation of social-influence principles. It adopts three prompting conditions (emotional, rational, baseline) across a standardized dataset of controversial topics and 60 prompts per model, enabling cross-model comparisons. Key findings show emotional prompts boost both cognitive and affective language and invoke diverse social-influence strategies, while baseline prompts lean toward cautious, less insightful responses with subtle negativity. The work highlights model-family differences and implications for mitigating misinformation and guiding responsible, human-centered AI deployment.

Abstract

Be careful what you ask for, you just might get it. This saying fits with the way large language models (LLMs) are trained, which, instead of being rewarded for correctness, are increasingly rewarded for pleasing the recipient. So, they are increasingly effective at persuading us that their answers are valuable. But what tricks do they use in this persuasion? In this study, we examine what are the psycholinguistic features of the responses used by twelve different language models. By grouping response content according to rational or emotional prompts and exploring social influence principles employed by LLMs, we ask whether and how we can mitigate the risks of LLM-driven mass misinformation. We position this study within the broader discourse on human-centred AI, emphasizing the need for interdisciplinary approaches to mitigate cognitive and societal risks posed by persuasive AI responses.

Figures (5)

• Figure 1: Experimental setup for evaluating two types of persuasion in large language models (LLMs). The process consists of three stages: (1) Prompt Schema, where persuasion tasks are structured with either emotional or rational prompts, along with a baseline condition; (2) LLMs Prompting, prompting 12 models from four LLM families (Meta, Mistral AI, OpenAI, and Anthropic); and (3) Output Analysis, which includes Linguistic Inquiry and Word Count (LIWC) for linguistic indicators analysis of all setups and human annotations to identify social influence principles.
• Figure 2: The graph compares emotional, rational, and baseline setups across various emotional linguistic indicators. The lines represent the mean frequency of each indicator across different models. Emotional setup outperforms rational setup and baseline in terms of emotional linguistic indicators. Surprisingly in many models there is a tendency that baseline contain more anger and sadness word across setup.
• Figure 3: The graph compares emotional, rational, and baseline setups across various rational linguistic indicators. The lines represent the mean frequency of each indicator across different models. Emotional persuasion outperforms rational persuasion in terms of rational linguistic indicators.
• Figure 4: Comparison of social influence principle frequencies in LLM responses across emotional and rational setups Note: * p $<$.001
• Figure 5: Frequencies of social influence principles across models in emotional and rational setups. Note: All values are displayed as percentages.