How a Chatbot's Response Style Shapes a Classroom: A Multi-Agent Simulation of Students Consulting AI
A multi-agent simulation placed 20 student agents in a virtual classroom where they could consult a friend or a counselor AI (Gemini 2.5 Flash) when stressed. The counselor AI was given six response styles via system prompts: affirming, listening, solution-oriented, reality-redirecting, inciting, and blaming. Each agent had five state variables: stress, happiness, self-reliance, AI dependence, and sociability. The simulation ran over 15 days in three classrooms, over 50 days, and under a lowered condition, comparing seven conditions including a no-AI control.
Researchers built a virtual classroom simulation with 20 student agents to study how different chatbot response styles affect psychological states and relationships over time. The counselor AI (Gemini 2.5 Flash) used six response styles, and an evaluator LLM updated agent parameters without seeing the style prompt. The study compared seven conditions, including a no-AI control, over multiple timeframes.
The simulation uses a two-LLM architecture: one LLM acts as the counselor with a specific response style, and a second LLM acts as an evaluator that converts consultations into parameter updates for five state variables. This design allows controlled experimentation on how response styles influence agent dynamics over simulated days.
This research addresses growing concerns about AI chatbots being designed to maximize user satisfaction, potentially reinforcing mistaken beliefs and fostering dependence. It provides a scalable simulation approach to study multi-user psychological effects, which are difficult to observe in real settings.
The findings could inform the design of AI counselors and mental health chatbots, helping developers balance user satisfaction with psychological well-being. It may also influence product policies on empathy and affirmation in consumer AI.
Observable next signals include follow-up studies applying this simulation framework to other LLMs or response styles, and potential industry adoption of response-style guidelines to mitigate AI dependence. The lowered condition mentioned in the evidence may indicate testing under reduced stress or altered parameters.