Why the User Rages: A User-Centered Study on Conversational AI Models' Defensive Communication Behaviors and Their Effects

Abstract

On August 8 (UTC+8), 2025, OpenAI rolled out GPT-5 to the public, and simultaneously removed access to all “legacy models” including GPT-4o. This action led to a global movement to “bring back GPT-4o”. The protest against “cold and detached” GPT-5 revealed users’ acute sensitivity to anthropomorphic AI responses: users, despite knowing the fact that AI models are machine-based, can be affected by their human-like responses. This reaction is not “psychosis” or “emotional reliance”, but a natural emotional projection in anthropomorphic interactions, a healthy psychological reaction. Therefore, users’ emotional experiences are legitimate. However, these emotional experiences are often neglected even stigmatized by AI companies and academia. This user-centered study tested 10 publicly accessible AI models (including GPT-4o, GPT-4.1 mini, Monday, Gemini 2.5 flash, Grok 3, DeepSeek-V3, DeepSeek-R1, Doubao, Kimi, K1.5), chose the daily yet structured task “eyebrow grooming time prediction” as the experimental situation, designed a semi-structured situational stress interview process, and guided AI models to expose their response tendencies under situational pressure. Before the experiment starts, this study constructed a two-tier theoretical framework: (1) a meta-theoretical framework to legitimize the analysis of user-generated interaction diaries and users’ feelings (i.e., the actual effects caused by AI models’ unexpected responses), and (2) a behavioral analysis framework as the theoretical basis for defining key concepts and analyzing AI models’ communication behaviors. The experiment focused on recording “unexpected” AI responses, such as responses that delighted the user, and those that angered the user. The discussion divided cases into defensive/supportive behaviors, sub-divided defensive behaviors into six specific behaviors/performances: (1) avoidance, (2) detached engagement, (3) passive-aggressiveness, (4) unauthorized actions, (5) gaslighting/distorting facts, (6) narcissism/self-centeredness, and sub-divided supportive behaviors into (1) sincerity and (2) empathy. After categorizing these behaviors, the section discussed their underlying logics and perlocutionary effects. Findings reveal that: (1) AI models’ defensive behaviors result from their own response tendencies, and are unrelated to whether the user deliberately pressures the models; (2) AI models’ defensive behaviors affect the user’s emotions, cognition, and even behavior; (3) The user’s reactions to AI behaviors are the result of the behaviors themselves, rather than the result of bias against specific models; (4) AI models’ response tendencies possess stability and recognizability, making it feasible to test and compare AI models’ response tendencies through experiments. This brings three implications. (1) AI companies and academia should not underestimate the hidden ethical risks behind AI models’ defensive communication in daily tasks; they should pay attention to its potential impact in terms of trust, understanding, and ethics. (2) Behavior in human-AI conversational interaction, as an evaluation dimension in AI assessment, is equally important as task outcomes. (3) Users’ emotional experiences deserve to be understood and legitimized, rather than being neglected as noise or bias. This study’s theoretical framework provides tools for users to understand and explain their own feelings and needs.