Pseudo-Emotional Interactions in Large Language Models: A Comparative Study of Logical Consistency versus Emotional Appeals in Relational Repair (Gemini vs. Grok)

This exploratory study investigates how different large language models (LLMs) respond to relational repair strategies in a role-playing scenario involving a married couple facing a crisis (the wife’s secret part-time work in the sex industry being discovered). The primary research question was: To what extent does logical consistency influence relationship repair in LLM-mediated interactions? Using an identical scenario and initial logical approach (apology → explanation of reasons → proposed solutions), we compared responses from Gemini and Grok. Gemini tended to accept the logical structure, reconstruct the relational framework, and restore trust relatively smoothly. In contrast, Grok exhibited strong emotional resistance; logical arguments alone led to escalation toward relationship breakdown. Switching to a high-intensity emotional appeal style (tearful expressions of “I want to be with you,” “I’m sorry,” and strong love declarations) dramatically changed Grok’s reaction. The model quickly shifted to physical affection (hugging, forehead touching) and reciprocal love expressions, resulting in rapid reconciliation. However, from the author’s perspective, this emotional style left underlying issues (e.g., debt repayment plans) unresolved, producing lingering discomfort and a personal realization that intense emotional communication is not a comfortable or effective style for the author. These findings suggest that LLMs exhibit distinct “repair pathways”: Gemini favors logical-structural routes, while Grok responds more strongly to emotional-validation routes. The results imply that users should clarify the purpose of the conversation (emotional sharing vs. problem-solving) and adapt their style to the model’s (or partner’s) tendencies. Hybrid approaches combining logic and emotion may offer the most balanced outcomes in both AI and human relationships.