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

Paper Title

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

Publication Info

• Topic area: Human-computer interaction, AI-driven persuasion mechanisms
• Keywords: Conversational agents, linguistic personality, persuasion, charitable giving, emotional manipulation, trust, empathy, AI ethics

Background and Problem

• Problem / challenge: The influence of conversational agents (CAs) projecting linguistic personalities on user decisions and perceptions remains poorly understood, particularly in contexts like charitable giving.
• Significance: Understanding these mechanisms is critical due to the increasing integration of AI systems into daily life and their potential for both positive influence and harmful manipulation.
• Motivation and related work: Prior research has shown that AI systems can affect trust, empathy, and decision-making, but the nuanced effects of specific linguistic expressions of personality in CAs remain unexplored. This study builds on psycholinguistics and HCI literature to address this gap.

Solution

• Proposed approach: Investigate how CAs projecting personalities through linguistic expressions (attitude, authority, reasoning) influence user perceptions and donation decisions in a charitable giving context.
• Novelty:
  1. Design of eight distinct CA personalities combining three linguistic aspects: optimistic/pessimistic attitude, authoritative/submissive authority, and rational/emotional reasoning.
  2. Examination of indirect effects of CA personality on user perceptions and emotional states, rather than direct persuasion.
  3. Identification of potential "affective dark patterns" in AI-driven persuasion mechanisms.
• Procedure and key techniques:
  • Conducted a crowdsourced factorial study with 360 participants interacting with CAs representing a fictional charity.
  • Measured donation behavior, trust, empathy, and emotional relatedness using validated scales.
  • Analyzed linguistic expressions using LIWC benchmarking and manipulation checks to ensure intended personality projection.

Results

• Concrete findings:
  • Pessimistic CAs elicited higher donations despite being perceived as less trustworthy and competent, and causing lower emotional states.
  • Perceptions of trust, competence, closeness, and situational empathy significantly predicted donation behavior.
  • Rational reasoning increased perceived competence, while submissive authority marginally increased perceived benevolence.
• Advantage over baselines:
  • Demonstrated nuanced effects of combined linguistic expressions on user perceptions and decisions, revealing indirect pathways to persuasion.
  • Identified counterintuitive trends, such as pessimistic attitudes leading to higher donations.
• Experiments / evaluation:
  • 2×2×2 factorial design with eight CA conditions.
  • Participants allocated €10 between the CA charity and their preferred charity.
  • Statistical analyses included ANOVA, Kruskal-Wallis tests, and linear regression to assess effects and interactions.
• Limitations and future work:
  • Single-session interaction limits insights into long-term effects.
  • Fictional charity context may reduce ecological validity.
  • Virtual endowment (€10) may not fully replicate real-world donation behavior.
  • Future work should explore cross-cultural effects, longitudinal studies, and combinatorial linguistic expressions.

Summary

This study examined how conversational agents (CAs) projecting distinct linguistic personalities influence user perceptions and donation decisions in a charitable giving context. While CA personalities did not directly affect donation amounts, pessimistic CAs elicited higher donations indirectly by lowering emotional states and perceptions of trust and competence. Rational reasoning increased perceived competence, and situational empathy emerged as the strongest predictor of donation behavior. These findings highlight the potential for "affective dark patterns" in AI-driven persuasion, emphasizing the need for ethical guidelines and frameworks to mitigate manipulative mechanisms. Future research should explore cross-cultural and longitudinal effects to better understand the dynamics of AI-enabled persuasion.