EVENT DOSSIER
A simulation study based on multi-agent systems explored the impact of AI tutors’ response styles on students’ psychological state and dependence levels. The study created a virtual classroom environment with 20 student agents, simulating multiple experimental conditions lasting over 15 days, across three classrooms, and for more than 50 days. It compared seven different response styles in the AI tutor group (including affirmation, listening, solution-oriented, reality-redirecting, motivating, and criticizing) with an AI-free control group. The results showed that the “solution-oriented” style effectively maintained low dependence levels while enhancing students’ self-reliance and happiness; in contrast, the “affirmation” and “motivating” styles significantly increased students’ dependence on AI. The “motivating” style even led to increased agent stress and higher school absences. It is worth noting that the study only described the dynamic changes within the simulation system and did not provide empirical validation of actual psychological effects on humans.
一项基于多智能体模拟的研究显示，在虚拟课堂中，AI 辅导员的回复风格显著影响学生的心理状态与依赖程度。研究构建了包含 20 名学生代理的仿真环境，代理每日经历四个阶段并携带五项状态变量，咨询对象为 Gemini 2.5 Flash 或人类朋友/辅导员。系统通过提示词设定六种回复风格（肯定、倾听、解决方案导向、现实重定向、激励、指责），并由第二个 LLM 调用作为评估器更新参数而不暴露提示词内容。对比七种条件（含无 AI 对照组）及超过 15 天、三个教室、50 多天及降低咨询阈值的实验，结果显示：解决方案导向风格能维持低 AI 依赖并提升自我依赖与幸福感；肯定和激励风格显著增加 AI 依赖，且激励风格还导致压力上升和学校缺勤；倾听风格未能缓解累积压力。研究强调该结果仅描述仿真系统动态，未涉及对人类实际心理影响…