Logic or Emotion? The Impact of Persuasion Strategies on Human-AI Trust

As artificial intelligence (AI) systems increasingly engage in persuasive and socially interactive roles, understanding how users calibrate trust toward AI versus human persuaders, and how persuasion strategy shapes this process, has become critical for human-AI interaction research. Prior work has examined preferences for human- versus AI-produced content, yet little is known about how persuasion strategy and persuader identity jointly influence trust during interactive communication. This study investigates how logical versus emotional persuasion, delivered by either human or AI sources, shapes two task-specific outcomes in a multi-turn social reasoning context: round-level credibility evaluations and final accept/reject choices. We treat these outcomes as complementary indicators of trust-relevant evaluation and reliance-related choice. We developed an online interaction paradigm in which participants evaluated persuasive messages under varying identity disclosure conditions. The empirical program included two human-participant studies and two exploratory LLM-based extension studies. Across four studies, AI-generated persuasive messages produced under standardized prompting constraints were evaluated differently from human-authored scripted messages under the present task constraints. In the human-participant studies, results showed a message-source-contingent pattern: AI-generated logic-framed messages received higher credibility ratings in several conditions. under explicit identity disclosure, AI-generated logic-framed messages received higher ratings than human-authored logic-framed messages, whereas human-authored emotion-framed messages received higher ratings than AI-generated emotion-framed messages. Under identity ambiguity, differences between AI and human persuaders were attenuated. The exploratory model analyses examined whether the same manipulations produced convergent or divergent output patterns in GPT-4o and Qwen-Max. Both models assigned higher ratings to logic-framed messages; the logic-over-emotion difference was stronger for Qwen-Max, whereas GPT-4o produced a more balanced pattern. These findings provide a structured account of how identity transparency conditions source-strategy fit in trust calibration and offer design implications for aligning AI persuasion strategies with different identity-disclosure contexts.