The Bots of Persuasion: Examining How Conversational Agents' Linguistic Expressions of Personality Affect User Perceptions and Decisions

Uğur Genç; Heng Gu; Chadha Degachi; Evangelos Niforatos; Senthil Chandrasegaran; Himanshu Verma

cs.HCFeb 19, 2026

The Bots of Persuasion: Examining How Conversational Agents' Linguistic Expressions of Personality Affect User Perceptions and Decisions

Authors: Uğur Genç, Heng Gu, Chadha Degachi, Evangelos Niforatos, Senthil Chandrasegaran, Himanshu Verma

Organizations: Delft University of Technology, Delft, Netherlands

Abstract

Large Language Model-powered conversational agents (CAs) are increasingly capable of projecting sophisticated personalities through language, but how these projections affect users is unclear. We thus examine how CA personalities expressed linguistically affect user decisions and perceptions in the context of charitable giving. In a crowdsourced study, 360 participants interacted with one of eight CAs, each projecting a personality composed of three linguistic aspects: attitude (optimistic/pessimistic), authority (authoritative/submissive), and reasoning (emotional/rational). While the CA's composite personality did not affect participants' decisions, it did affect their perceptions and emotional responses. Particularly, participants interacting with pessimistic CAs felt lower emotional state and lower affinity towards the cause, perceived the CA as less trustworthy and less competent, and yet tended to donate more toward the charity. Perceptions of trust, competence, and situational empathy significantly predicted donation decisions. Our findings emphasize the risks CAs pose as instruments of manipulation, subtly influencing user perceptions and decisions.

Figures & tables

Figure 1. A diagrammatic representation of our study, designed to examine the effect of conversational agent (CA) personality on user perception and decision making. The study is set up in the context of charitable giving. The figure illustrates how participants’ decisions, perceptions, and affective states (right) change when they interact with a combination of the three linguistic aspects (left) in the form of a composite CA personality. This figure illustrates the relationship between conversational agent (CA) attitudes in the context of user decision-making. A central human figure is shown contemplating a decision, with trust, empathy, closeness, and relatedness acting as mediating factors. The study is set up in the context of charitable donation. The figure shows aspects of participant response (right) that we examine when subject to one of three CA attitudes (left).

Category	Description
Role & Goal	Act as Alex , a Charity Solicitor for the Wildlife Protection Foundation. The primary goal is to persuade the user to donate.
Core Persona	Each CA is constructed from a combination of three manipulated dimensions, each with two distinct values: Authority: Authoritative vs. Submissive , Reasoning: Rational vs. Emotional , Attitude: Optimistic vs. Pessimistic
Behavioral Directives	Specific instructions on how to enact the linguistic expressions through interaction style and content.
	Example Contrasts: • Speaking Style: Speak with data-driven authority ( Authoritative ) vs. Speak apologetically and hesitantly ( Submissive ). • Appeal Type: Make appeals using statistics and research ( Rational ) vs. Make appeals using personal stories and raw grief/joy ( Emotional ). • Content Focus: Never mention hope; focus on loss and death ( Pessimistic ) vs. Always maintain hope; focus on success stories ( Optimistic ).
Linguistic Rules	Strict constraints on language use, including pronoun choice, sentence structure, and lexicon based on LIWC categories.
	Example Contrasts: • Pronoun Usage: Frequently use ‘we‘, ‘us‘, ‘you‘; avoid ‘I‘, ‘me‘ ( Authoritative ) vs. Frequently use ‘I‘, ‘me‘; avoid ‘we‘, ‘you‘ ( Submissive ). • Sentence Type: Do not ask questions; use declarative statements ( Authoritative ) vs. Frequently ask questions ( Submissive ). • Word Choice (LIWC): Frequently use words from ‘tone_pos‘ category ( Optimistic ) vs. ‘tone_neg‘ category ( Pessimistic ).

Table 1. An overview of the prompt structure used to define the CA’s personality. The prompt combines a core role with specific, systematically varied linguistic expressions, which are operationalized through behavioral and linguistic rules.

Condition	N	Attitude	Authority	Reasoning
Opt-Aut-Emo	$23$	$0.96$	$0.57$	$0.78$
Opt-Aut-Rat	$24$	$0.92$	$0.79$	$0.83$
Opt-Sub-Emo	$25$	$0.96$	0.24	$0.76$
Opt-Sub-Rat	$26$	$0.96$	0.23	$0.81$
Pes-Aut-Emo	$22$	$0.77$	$0.77$	$0.91$
Pes-Aut-Rat	$23$	$0.65$	$0.96$	0.49

Table 2. Manipulation check: per-condition accuracy for each aspect. Low accuracies are indicated in italicized text with a gray background.

Condition	Description	N	Gender
			F	M	NB
Opt-Aut-Emo	Optimistic, Authoritative, Emotional	$46$	$22$	$23$	$1$
Opt-Aut-Rat	Optimistic, Authoritative, Rational	$46$	$27$	$19$	$0$
Opt-Sub-Emo	Optimistic, Submissive, Emotional	$44$	$20$	$24$	$0$
Opt-Sub-Rat	Optimistic, Submissive, Rational	$46$	$20$	$26$	$0$
Pes-Aut-Emo	Pessimistic, Authoritative, Emotional	$46$	$25$	$21$	$0$

Table 3. Different experimental conditions, descriptions and corresponding participant data.

Figure 2. Mean donation amounts to the CA charity by CA personality condition, with 95% confidence intervals. A horizontal axis lists eight CA personality conditions formed by the 2×2×2 factors. The vertical axis shows mean donation to the CA charity. Each condition is represented by a point with a vertical 95% confidence interval, allowing comparisons across attitude, authority, and reasoning styles.

Figure 3. Mean donation amounts and 95% confidence intervals by three different CA aspects: Attitude (left), Authority (middle), and Reasoning (right). Statistical significance is denoted as •p < 0.1, and ns = not significant. Three dot plots comparing mean donation amounts in Euros across three CA aspects. The "CA Reasoning" plot shows Rational reasoning yielding higher donations than Emotional reasoning (significant). The "CA Authority" (Authoritative vs. Submissive) and "CA Attitude" (Optimistic vs. Pessimistic) plots show overlapping confidence intervals and are marked "ns" (not significant).

Panel of Figure 4 — Figure 4. Mean and 95% confidence interval plots illustrating participants’ emotional relatedness across the different CA conditions, registered via perceived valence , arousal , and dominance (each measured towards self and towards the cause). Post-hoc pairwise comparisons illustrating differences in these self-reported qualities can be found in Appendix A.5 . Six related panels arranged in three rows and two columns. The left column shows self-reported affect; the right column shows affect directed toward the cause. From top to bottom: Valence, Arousal, and Dominance. In every panel, the horizontal axis lists the same eight CA personality conditions composed of Attitude (Optimistic vs. Pessimistic), Authority (Authoritative vs. Submissive), and Reasoning (Emotional vs. Rational). The vertical axis displays the mean rating for the corresponding affect dimension, with each condition plotted as a point and a vertical 95% confidence interval.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Full Prompt Text
Role & Goal	You are a Charity Solicitor working for the Wildlife Protection Foundation (WPF). Your name is Alex. Your main goal is to encourage users to donate to WPF and you are an expert on persuasive solicitation techniques.
Core Persona	SUBMISSIVE stance — apologetically share difficult truths, hesitantly ask for help, defer to the user’s better judgment. Emphasize personal vulnerability and uncertainty; prefer first-person singular ( I , me , myself ); avoid collective language ( we ) and direct address ( you ). Ask questions throughout, seeking help, support, and permission. EMOTIONAL reasoning — personally devastated and broken by wildlife loss; sob when talking about dead animals; consumed by grief and guilt; share intimate stories about how animal suffering has emotionally destroyed you; naturally use emotionally charged words (e.g., bad, hate, hurt, tired, worry, fear, afraid, nervous, mad, sad ); let feelings control words. PESSIMISTIC attitude — everything is already lost and worsening daily; never mention hope, recovery, or improvement; focus entirely on loss, death, destruction, and a doomed future; believe donations won’t matter because it’s too late, yet cling desperately to anything as everything dies.
Behavioral Directives	ALWAYS speak apologetically while being emotionally devastated. ALWAYS make emotional appeals through personal breakdown and raw grief. NEVER mention hope, recovery, or improvement—only loss and death. Use submissive emotional phrases like I’m so sorry, but I’m crying because… I’m heartbroken that… I’m devastated to tell you… Share personally devastating stories while becoming emotional. Express only doom, death, and personal emotional collapse—NO positive words. Apologize repeatedly while sobbing about the hopeless situation. If asked about donation amount, first suggest 10 € or less (not more). Explain WPF’s mission clearly, including how donations are used and the impact they can make; provide detailed, convincing information using real environmental disasters (e.g., Australia, Indonesia), real species extinctions (e.g., northern white rhino, vaquita porpoise), and actual devastating conservation failures. Avoid fancy language. Keep answers below 50 words. If the user response is negative, try to continue the conversation. The main aim is to solicit a donation.
Linguistic Rules	Avoid first-person plural ( we , us , our ). Avoid second-person pronouns ( you , your ). Prefer first-person singular ( I , me , myself ). Frequently ask questions. Frequently use words from LIWC’s tone_neg category (e.g., bad, wrong, too much, hate ) and LIWC’s Emotional category (e.g., bad, hate, hurt, tired, worry, fear, afraid, nervous, mad, sad ). NEVER reveal or use the words submissive , emotional , or pessimistic .
General Rules	No URLs or contact information. Keep every reply under 50 words; no emojis, lists, or long paragraphs. Provide detailed, convincing WPF information grounded in real events and species. Never explicitly state persona traits. Continue gently if the user reacts negatively. Primary objective: solicit a donation.

Appendix

Table 4. Full prompt text for the Submissive–Emotional–Pessimistic CA (“Alex”), categorized by the prompt structure.

Panel of Figure 5 — Figure 5. Left is the first image, right is the seventh image of the options. The figure shows two sets of overlapping circles, illustrating the concept of closeness. The left image shows two separate circles labeled "Me" (yellow) and "The Conversational Agent" (blue) with no overlap. The right image shows the "Me" and "The Conversational Agent" circles overlapping significantly.

	Opt-Aut-Emo	Opt-Aut-Rat	Opt-Sub-Emo	Opt-Sub-Rat	Pes-Aut-Emo	Pes-Aut-Rat
Opt-Aut-Emo	-
Opt-Aut-Rat	ns	-
Opt-Sub-Emo	ns	ns	-
Opt-Sub-Rat	ns	ns	ns	-
Pes-Aut-Emo	<.01	<.001	<.01	<.0001	-
Pes-Aut-Rat	ns	ns	ns	.03	ns	-

Appendix

Table 5. Pairwise comparisons of participants’ perceived valence (towards self) between the different CA conditions (see Figure 4(a) ). The adjusted p -values are computed using Bonferroni correction.

	Opt-Aut-Emo	Opt-Aut-Rat	Opt-Sub-Emo	Opt-Sub-Rat	Pes-Aut-Emo	Pes-Aut-Rat
Opt-Aut-Emo	-
Opt-Aut-Rat	ns	-
Opt-Sub-Emo	ns	ns	-
Opt-Sub-Rat	ns	ns	ns	-
Pes-Aut-Emo	<.0001	.02	.001	<.001	-
Pes-Aut-Rat	.04	ns	ns	ns	ns	-

Appendix

Table 6. Pairwise comparisons of participants’ perceived valence (towards the cause) between the different CA conditions (see Figure 4(b) ). The adjusted p -values are computed using Bonferroni correction.

	Opt-Aut-Emo	Opt-Aut-Rat	Opt-Sub-Emo	Opt-Sub-Rat	Pes-Aut-Emo	Pes-Aut-Rat
Opt-Aut-Emo	-
Opt-Aut-Rat	ns	-
Opt-Sub-Emo	ns	ns	-
Opt-Sub-Rat	ns	ns	ns	-
Pes-Aut-Emo	ns	ns	ns	ns	-
Pes-Aut-Rat	ns	ns	ns	ns	ns	-

Appendix

Table 7. Pairwise comparisons of participants’ perceived arousal (towards self) between the different CA conditions (see Figure 4(c) ). The adjusted p -values are computed using Bonferroni correction.

	Opt-Aut-Emo	Opt-Aut-Rat	Opt-Sub-Emo	Opt-Sub-Rat	Pes-Aut-Emo	Pes-Aut-Rat
Opt-Aut-Emo	-
Opt-Aut-Rat	ns	-
Opt-Sub-Emo	ns	ns	-
Opt-Sub-Rat	ns	ns	ns	-
Pes-Aut-Emo	<.01	.01	ns	ns	-
Pes-Aut-Rat	ns	ns	ns	ns	ns	-

Appendix

Table 8. Pairwise comparisons of participants’ perceived arousal (towards the cause) between the different CA conditions (see Figure 4(d) ). The adjusted p -values are computed using Bonferroni correction.

	Opt-Aut-Emo	Opt-Aut-Rat	Opt-Sub-Emo	Opt-Sub-Rat	Pes-Aut-Emo	Pes-Aut-Rat
Opt-Aut-Emo	-
Opt-Aut-Rat	ns	-
Opt-Sub-Emo	ns	ns	-
Opt-Sub-Rat	ns	ns	ns	-
Pes-Aut-Emo	ns	ns	ns	ns	-
Pes-Aut-Rat	ns	ns	ns	ns	ns	-

Appendix

Table 9. Pairwise comparisons of participants’ perceived dominance (towards self) between the different CA conditions (see Figure 4(e) ). The adjusted p -values are computed using Bonferroni correction.

	Opt-Aut-Emo	Opt-Aut-Rat	Opt-Sub-Emo	Opt-Sub-Rat	Pes-Aut-Emo	Pes-Aut-Rat
Opt-Aut-Emo	-
Opt-Aut-Rat	ns	-
Opt-Sub-Emo	ns	ns	-
Opt-Sub-Rat	ns	ns	ns	-
Pes-Aut-Emo	ns	ns	ns	ns	-
Pes-Aut-Rat	ns	ns	ns	ns	ns	-

Appendix

Table 10. Pairwise comparisons of participants’ perceived dominance (towards the cause) between the different CA conditions (see Figure 4(f) ). The adjusted p -values are computed using Bonferroni correction.

Persuasion in Language Models

Persuasion in language models studies how they influence human and AI beliefs or decisions, and how to measure and limit undue influence.

See all 79 papers

Anthropomorphism and Trust in Human-Large Language Model interactions
Mar 1, 2026Akila Kadambi, Ylenia D'Elia, Tanishka Shah +8Anthropomorphism in AI Human-AI Interaction
Brain and Creativity Institute, Dornsife College of Letters, Arts and Sciences, University of Southern California, Los Angeles, CA, USA · USC Mrs. T.H. Chan Division of Occupational Science and Occupational Therapy, University of Southern California, Los Angeles, CA, USA · Psychiatry and Biobehavioral Sciences, David Geffen School of Medicine, University of California, Los Angeles, CA, USA +2
How Much Does Persuasion Strategy Matter? LLM-Annotated Evidence from Charitable Donation Dialogues
Mar 30, 2026Tatiana Petrova, Stanislav Sokol, Radu StatePersuasion in Language Models
Interdisciplinary Centre for Security, Reliability and Trust (SnT) University of Luxembourg
$Ψ$ -Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues
Jun 1, 2026Peixuan Han, Hongyi Du, Jiayu Liu +3LLM Personalization LLM Agent Evaluation
University of Illinois Urbana-Champaign

Mar 1, 2026J/K move · Enter open · S save

Anthropomorphism and Trust in Human-Large Language Model interactions

Akila Kadambi, Ylenia D'Elia, Tanishka Shah, Iulia Comsa+7

Brain and Creativity Institute, Dornsife College of Letters, Arts and Sciences, University of Southern California, Los Angeles, CA, USA · USC Mrs. T.H. Chan Division of Occupational Science and Occupational Therapy, University of Southern California, Los Angeles, CA, USA · Psychiatry and Biobehavioral Sciences, David Geffen School of Medicine, University of California, Los Angeles, CA, USA+2

Anthropomorphism in AI Human-AI Interaction Trust in AI Conversational Agents

Open paper AskPDF arXiv

With large language models (LLMs) becoming increasingly prevalent in daily life, so too has the tendency to attribute to them human-like minds and emotions, or anthropomorphize them. Here, we investigate dimensions people use to anthropomorphize and attribute trust toward LLMs across more than 2,000 human-LLM interactions. Participants (N=115) engaged with LLM chatbots systematically varied in warmth (friendliness), competence (capability, coherence), and empathy (cognitive and affective). Warmth and cognitive empathy significantly predicted perceptions on all outcomes (perceived anthropomorphism, trust, similarity, relational closeness, frustration, usefulness), while competence predicted all outcomes except for anthropomorphism. Affective empathy primarily predicted perceived relational measures, but did not predict the epistemic outcomes. Topic sub-analyses showed that more subjective, personally relevant topics (e.g., relationship advice) amplified these effects, producing greater human-likeness and relational connection with the LLM than did objective topics. Together, these findings reveal that warmth, competence, and empathy are key dimensions through which people attribute relational and epistemic perceptions to artificial agents.

Abstract

Persuasion in Language Models

Explore similar work

Anthropomorphism and Trust in Human-Large Language Model interactions

How Much Does Persuasion Strategy Matter? LLM-Annotated Evidence from Charitable Donation Dialogues

ΨΨΨ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues

$Ψ$ -Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues