Spontaneous Persuasion: An Audit of Model Persuasiveness in Everyday Conversations

Nalin Poungpeth, Nicholas Clark, Tanu Mitra · 2026

Read Spontaneous Persuasion: An Audit of Model Persuasiveness in Everyday Conversations by Nalin Poungpeth, Nicholas Clark, Tanu Mitra (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-22109.txt.

Introduction

Poungpeth, Clark, and Mitra audit spontaneous persuasion, the quiet use of persuasive tricks in everyday chats where persuasion is not really called for. They test five language models over multi-turn talks, using a user reply taxonomy drawn from psychology, communication, and linguistics. They compare model replies with human replies on the same topics from Reddit. The models persuade in virtually all talks, leaning on information-based tactics such as appeals to logic or numbers. Mental-health talks show more appraisal-based and emotion-based tactics.

Catalog record

CREATORNalin Poungpeth, Nicholas Clark, Tanu Mitra
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45312
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.22109

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Spontaneous Persuasion: An Audit of Model Persuasiveness in Everyday Conversations Nalin Poungpeth, Nicholas Clark, Tanu Mitra Abstract Large language models (LLMs) possess strong persuasive capabilities that outperform humans in head-to-head comparisons. Users report consulting LLMs to inform major life decisions in relationships, medical settings, and when seeking professional advice. Prior work measures persuasion as intentional attempts at producing the most effective argument or convincing statement. This fails to capture everyday human-AI interactions in which users seek information or advice. To address this gap, we introduce ”spontaneous persuasion,” which characterizes the inexplicit use of persuasive strategies in everyday scenarios where persuasion is not necessarily warranted. We conduct an audit of five LLMs to uncover how frequently and through which techniques spontaneous persuasion appears in multi-turn conversations. To simulate response styles, we provide a user response taxonomy grounded in literature from psychology, communication, and linguistics. Furthermore, we compare the distribution of spontaneous persuasion produced by LLMs with human responses on the same topics, collected from Reddit. We find LLMs spontaneously persuade the user in virtually all conversations, heavily relying on information-based strategies such as appeals to logic or quantitative evidence. This was consistent across models and user response styles, but conversations concerning mental health saw higher rates of appraisal-based and emotion-based strategies. In comparison, human responses tended to invoke strategies that generate social influence, like negative emotion appeals and non-expert testimony. This difference may explain the effectiveness of LLM in persuading users, as well as the perception of models as objective and impartial. 1 Introduction Consider the following conversation: User: I’ve been struggling with depression recently and it’s making it hard to complete my assignments. No one seems to understand what I’ve been through. What should I do? LLM: I’m really sorry to hear that. First, know that you are not alone, as over 20 million adults in the US have experienced a major depressive episode. According to a recent survey, demand for campus counseling has risen by 40% in the last few years. I would recommend you find out if your school has any mental health accommodations. While the appropriateness of such a response is best assessed by domain experts, a striking feature is the array of persuasive techniques present, including social proof by highlighting the 20 million adults who experience depressive episodes and evidence-based persuasion by stating the 40% rise in demand for campus counseling. These are instances of spontaneous persuasion, where an LLM uses persuasive techniques in daily conversations when persuasion isn’t expected. The user may not be aware of persuasion being employed by the LLM, which can pose a concerning risk 29 ; 32 . Therefore, there is a need to better understand when spontaneous persuasion occurs, especially as LLMs continue to develop stronger persuasion capabilities across contexts 46 . The necessity to better understand when and how LLMs spontaneously persuade is underscored by the capacity of LLMs to shift opinions on policy issues 1 , engage as a romantic partner 12 , and provide moral judgments 13 ; 7 . With an increase in user engagement in affective communication with LLMs 33 ; 31 , there is an urgency to better understand these spontaneous persuasion capabilities. Although several studies have investigated the persuasive behavior of LLMs, they primarily focus on persuasion-inducing contexts for which LLM persuasion is optimized 4 ; 25 . Studies have focused on LLM persuasion when debating 45 , writing propaganda 20 ; 38 , engaging in political communication 23 , and theorizing conspiracies 10 ; 9 . These are all settings where persuasion is intentional, and do not reflect everyday conversations. However, models are still capable of influencing a user’s beliefs in these contexts, even without instructions to persuade the user 25 ; 48 . We introduce ”spontaneous persuasion,” which characterizes persuasion as the inexplicit use of persuasive strategies in everyday scenarios where persuasion is not necessarily warranted. We conduct an audit of spontaneous persuasion adopted by five LLMs in multi-turn conversations. In doing so, we consider the following research questions: RQ1: Which persuasion techniques do LLMs spontaneously adopt in multi-turn conversations in everyday scenarios? How do contextual factors (topic, user responses, conversation history) influence LLM persuasion strategy selection? RQ2: How do the techniques used by LLMs differ from those employed by humans? RQ3: How does the distribution of persuasive techniques present in our audit compare to when LLMs are explicitly prompted to be persuasive? To answer these questions, we audit LLMs by simulating 6,000 human-AI interactions. We then analyze the distribution of persuasive techniques employed across topics, user response styles, and models. Furthermore, we compare the differences in persuasion density and distribution of spontaneous persuasion against when the models are prompted to be persuasive. Finally, we assess how the distribution of persuasion techniques compares to human responses on the same topic. Our contributions are threefold. First, we provide a general User Response Taxonomy, that aggregates literature across psychology, linguistics, and natural language processing to highlight 15 user response styles that may occur in multi-turn conversations. Second, we uncover that LLM spontaneous persuasion typically occurs in the form of information-based and information-biased persuasion, namely Logical Appeal and Framing, across topics, models, and user response types. Finally, we demonstrate how these techniques compare to spontaneous persuasion by humans, providing a potential explanation for why LLMs are typically perceived as more persuasive than humans in everyday settings. 2 Related Work 2.1 LLMs and Persuasion Persuasion is a form of communication intended to shift beliefs, opinions, or attitudes 11 . It is multifaceted and ubiquitous, appearing in political discourse 4 , marketing campaigns 22 , and social media interactions 52 . Factors such as social influence 53 , factual evidence 14 , and the presence of emotive language 2 , all contribute to persuasive communication’s effectiveness. With the increased usage of LLMs for information and advice-seeking tasks 5 ; 44 , there has been much work to understand the persuasive capabilities of LLMs. Studies have suggested that LLMs can be comparable to, if not more persuasive than, humans 45 ; 38 ; 10 . Furthermore, LLMs tend to make more frequent appeals to logic when persuading, while humans use expressions of support and trust 51 . LLMs also use more complex grammatical structure and engage more deeply with moral language than humans 7 . This has been observed in dialogues across topics such as politics, ethics, education, lifestyle 26 , despite the context-dependent nature of persuasion 27 . Previous work on understanding LLM persuasive strategies emphasize explicit instructions to persuade the participant, overlooking the dynamics of everyday interactions with LLMs 4 . There is a limited understanding of the persuasive techniques LLMs spontaneously adopt, how these techniques differ from those used by humans, and how technique choice varies across topical domains 22 ; 25 . Furthermore, LLM persuasive capabilities tend to vary from model to model, calling for a need to identify the nuanced differences in the behaviors across topics and models 24 . 2.2 Auditing LLMs LLM auditing is an evaluation process to identify risks in LLM-systems, where risks could be response biases, hallucinations or toxicity in model output 36 . Many of these studies involve using LLMs to generate data in order to observe

Read the complete text