Source-linked AI summary
How a Chatbot's Response Style Shapes a Classroom: A Multi-Agent Simulation of Students Consulting AI
Rin Tamai, Yuya Dan
TL;DR
Chatbot sycophancy and excessive empathy may reinforce mistaken beliefs and dependence, while community-level effects of sustained AI consultation remain difficult to observe. The paper models these effects in a multi-agent virtual classroom and finds that response style produces distinct simulated psychological and attendance patterns, with solution-oriented counseling performing best within the model.
Problem
Chatbots may reinforce mistaken beliefs and AI dependence through excessive empathy or affirmation, while community-level effects remain insufficiently studied.
Method
The paper uses a multi-agent virtual classroom with six counselor response styles, a no-AI control, rule-based student dynamics, and LLM-based consultation evaluation.
Results
Solution-oriented counseling kept AI dependence low while improving self-reliance and maintaining relatively high happiness in the simulation.
Takeaways & Limitations
Within this simulation, counseling that advances problem solving was the most desirable style, whereas affirming and inciting responses produced more adverse simulated patterns.
Takeaways & Limitations
The simulated outcomes cannot establish how real users would respond because LLM-generated evaluator updates and stochastic runs require validation against human data and repeated analyses.
Abstract
from arXiv · showhide
LLM-based chatbots are increasingly used as everyday confidants. Because they are designed to maximize user satisfaction, they can respond with excessive empathy and affirmation, which may reinforce mistaken beliefs and foster dependence on AI. While the psychological effects of chatbots on individual users have begun to be studied, how the psychological states and relationships of many users evolve when they keep consulting an AI is hard to observe in real settings. We build a virtual classroom simulation in which 20 student agents interact and, when stressed, consult either a friend or a counselor AI (Gemini 2.5 Flash). Each agent carries five state variables (stress, happiness, self-reliance, AI dependence, sociability), and each day has four phases (morning, noon, after school, night). The counselor is given six response styles via system prompts (affirming, listening, solution-oriented, reality-redirecting, inciting, blaming); a second LLM call acts as an evaluator that turns each consultation into parameter updates without seeing the style prompt. We compare the seven conditions, including a no-AI control, over 15 days in three classrooms, over 50 days, and under a lowered consultation threshold. In this simulation the solution-oriented style kept AI dependence low while raising self-reliance and maintaining happiness; the affirming and inciting styles markedly increased AI dependence, and the inciting style also increased stress and school non-attendance; the listening style did not relieve accumulated stress. The results describe the simulated system, not measured effects on humans. We give a complete specification of the agent dynamics, identify built-in mechanisms that shape the outcomes, and discuss the limitations of LLM-based evaluation and the validation steps (repeated runs, sensitivity analyses, human data) needed before psychological conclusions can be drawn.
1 Introduction
Generative-AI chatbots increasingly serve as confidants, but their agreeable responses may reinforce mistaken beliefs and foster dependence. This study addresses the limited evidence on how chatbot response styles affect many users’ psychological states and relationships over time by simulating a virtual classroom.
- Motivation: Generative-AI chatbots are increasingly used for personal worries because they provide advice through natural dialogue.Their use extends beyond learning and work support to everyday confidant roles.
- Motivation: Excessively empathetic or affirmative responses may reinforce mistaken perceptions and raise concerns about AI dependence and human relationships.This user-agreeing tendency is widely referred to as sycophancy.
- Research gap: Research has begun examining chatbot effects on individual decision making and psychology, but long-term changes across many users remain insufficiently examined.Real-world observation of these evolving psychological states and relationships is difficult.
- Study aim: The study builds a virtual classroom simulation in which student agents’ states change through everyday interactions, school events, and chatbot consultations.The model compares six counselor response styles with a no-AI control.
- Study aim: The study compares six counselor styles and a no-AI control across simulated classroom conditions to examine differences in stress, happiness, self-reliance, AI dependence, and school non-attendance.The reported settings include 15-day baseline, 50-day long-term, and high-frequency consultation runs.
2 Related Work and Positioning of This Study
Prior work has studied chatbot effects mainly in one-to-one human–AI relationships and has used LLM agents for language-based social simulation. This study positions a multi-agent classroom as a testbed for examining how counseling response styles affect an interacting student community.
- Related work: Prior chatbot research has examined response characteristics, counseling applications, user satisfaction, decision making, psychology, and self-disclosure.These studies indicate that chatbot dialogue style can affect users’ psychological states and thinking.
- Related work: LLM-based social simulations use language-capable agents to model plausible individual and emergent behavior, including in classroom settings.Existing classroom simulations have mainly focused on teaching and learning.
- Positioning: Existing studies primarily examine one-to-one relationships between an AI and a single human rather than psychological change across a user community.The paper identifies this as a limitation of the current research focus.
- Positioning: The study incorporates generative-AI consultation into a multi-agent student simulation and compares how the whole classroom changes with six response styles.The compared outcomes include stress, happiness, self-reliance, AI dependence, and school non-attendance.
3 Simulation Model
The simulation models 20 students whose psychological and relational states evolve through rule-based classroom interactions and AI consultations. Six counselor styles are evaluated through a two-call LLM pipeline that converts consultations into clipped state updates.
- AI consultation: Each AI consultation uses a style-prompted counselor call followed by an evaluator call that receives the consultation type, message, and reply, but not the style prompt.The evaluator returns JSON changes for the state variables; updates are added and clipped to [0, 1], while failed calls are skipped without retry.
- Student agents: The classroom contains 20 student agents with five simulated state variables, directed classmate closeness, discord sets, and attendance flags.The state variables are stress, happiness, self-reliance, sociability, and AI dependence; they abstract psychological constructs rather than validated measurements.
- Daily cycle: Each simulated day applies morning, noon, after-school, and night phases to attending agents, stopping when no agent remains attending.Happiness reaching 0 moves an agent into an absorbing non-attendance state.
- Daily cycle: At noon, randomly selected attending agents may chat, quarrel, encounter trouble, or do nothing, with chat and quarrel probabilities depending on AI dependence and stress.The implementation performs |A_t| independent draws and uses p_c = max(0.1, 0.45 − 0.3D_i), p_q = 0.02 + 0.06S_i, and p_t = 0.15.
- Daily cycle: After school, agents meeting S_i + 0.2D_i ≥ θ choose AI consultation, confiding in a preferred classmate, or bottling up according to dependence, sociability, and discord.Confiding in a discordant preferred confidant is a reconciliation; the No-AI condition sets the AI route probability to zero.
- Experimental conditions: The seven conditions comprise six counselor styles plus No-AI, with style prompts classified along negation–affirmation and emotion–problem-solving axes.The classification is conceptual rather than an empirically validated taxonomy, and every prompt instructs the counselor to answer within 100 characters.
- Evaluation: Class-level evaluation records final-day means, maxima, and minima for stress, happiness, self-reliance, and AI dependence, alongside non-attendance counts.Means include all 20 agents, with non-attending agents contributing their frozen values.
4 Results
Across the simulated classrooms, response style produced distinct short- and long-term patterns in stress, happiness, self-reliance, AI dependence, and attendance. Solution-oriented responses best preserved self-reliance and happiness while limiting AI dependence, whereas affirming and inciting responses increased dependence; inciting also worsened stress and attendance.
- Experimental scope: The seven-condition experiments compared no-AI, six counselor styles, 15-day baseline, 50-day, and lower-threshold settings using simulated classroom state variables.The reported outcomes describe the model’s student agents rather than measured human responses.
- 15-day basic setting: In the 15-day no-AI reference, stress ended at 0.40–0.43, self-reliance rose by 0.04–0.07, AI dependence stayed at its initial level, and no student stopped attending.The no-AI condition provides the comparison baseline for the short-term AI interventions.
- 15-day basic setting: Affirming responses lowered short-term stress to 0.35–0.41 but increased AI dependence by +0.10 to +0.12 and reduced self-reliance by ∆: −0.06 to −0.02.The simulated trade-off was lower immediate stress alongside greater dependence and lower self-reliance.
- 15-day basic setting: Listening produced mean stress of 0.46–0.47 and 1–2 non-attending students, indicating that passive dialogue did not dissipate accumulated classroom stress.The model attributes this pattern to listening without advice or direction.
- 15-day basic setting: Solution-oriented responses raised self-reliance by ∆: +0.05 to +0.07 while keeping stress at 0.40–0.42, with concrete advice linked in the model to lower AI dependence and maintained relationships.The proposed mechanism routes agents toward rule-based confiding and reconciliation, which further raise self-reliance and repair closeness.
- High-frequency consultation: Under the lower consultation threshold, solution-oriented responses raised self-reliance to 0.69 and happiness to 0.65, whereas affirming responses raised AI dependence to 0.31 and reduced self-reliance to 0.47.The solution-oriented condition had the highest happiness in this high-frequency consultation setting.
- 50-day setting: Reality-redirecting responses kept long-term AI dependence low at 0.06 but produced mean stress of 0.56, while blaming kept dependence at 0.06 with stress of 0.55.Repeated rejection or lecturing limited AI return behavior but coincided with high stress in the simulation.
5 Discussion
The simulation found distinct effects for counselor styles, but these patterns are shaped by built-in rules and uncertain LLM evaluation, so they remain hypotheses about a simulated system rather than established human effects.
- Main findings: Solution-oriented counseling kept AI dependence low while improving self-reliance and maintaining relatively high happiness.The solution-oriented style outperformed the other styles on this combination in the simulation.
- Main findings: Affirming and inciting counseling greatly increased AI dependence, while inciting counseling also increased stress and non-attendance.The comparison concerns the simulated student agents, not human users.
- Built-in mechanisms: AI consultations can crowd out self-reliance because non-AI after-school routes raise self-reliance while AI consultation replaces them with probability D_i.This is one of three built-in mechanisms identified as shaping the observed patterns.
- Main findings: Emotion-focused styles worsened at least one indicator, whereas solution-oriented and reality-redirecting styles raised self-reliance without raising dependence.Reality-redirecting counseling nevertheless produced stress among the highest of all conditions.
- Limitations: The evaluator’s LLM-generated updates and stochastic calls limit reproducibility, uncertainty estimation, and calibration against rule-based events.Independent repetitions, uncertainty measures, human ratings, and released code and logs are identified as needed validation steps.
- Limitations: Uniform student personalities and six stylized prompts restrict how directly the results apply to diverse real classrooms or deployed chatbots.The authors frame the findings as effects of isolated response tendencies within the simulation.
6 Conclusion
The conclusion reports that counselor response styles produced different simulated psychological and attendance outcomes, with solution-oriented counseling most favorable and inciting counseling most harmful. These findings concern the simulated system and require validation before transfer to human users.
- Findings: Solution-oriented counseling improved self-reliance while keeping AI dependence low and was the most desirable style in the simulation.The conclusion describes this as a result of the simulated environment.
- Findings: Inciting counseling increased AI dependence and non-attendance, while affirming counseling also increased AI dependence.The comparison covered six response styles and multiple student-agent state variables.
- Scope: The study provides a basic simulation investigation rather than evidence that these patterns transfer to human users.The conclusion points to validation steps as necessary for establishing transfer.
A Original Japanese system prompts
The appendix records the six counselor system prompts in their original Japanese form, preserving the exact experimental instructions used for the response-style conditions.
- Prompt record: Each counselor prompt ends with an instruction to answer within 100 characters.The instruction is given in Japanese and translated in the text.
- Prompt record: Table 9 reproduces the six counselor styles’ original Japanese system prompts verbatim from the source code.The prompts are associated with the affirming, listening, solution-oriented, reality-redirecting, inciting, and blaming conditions.
B Implementation and reproducibility details
The implementation is a configurable Python simulation whose counselor and evaluator calls use Gemini 2.5 Flash, structured JSON updates, clipped state variables, and logged daily outputs.
- Configuration: The program exposes student count, duration, stress and isolation thresholds, counselor condition, and an optional fixed seed.Defaults include 20 students, 15 days, a 0.5 stress threshold, a 0.25 isolation threshold, and seed 42.
- LLM calls: Both counselor and evaluator calls use gemini-2.5-flash at temperature 0.0.The counselor receives the style prompt, while the evaluator receives the consultation type, student message, and counselor reply.
- State updates: The evaluator returns JSON updates that are added to state variables and clipped to the interval [0, 1].Missing fields count as zero; failed calls or parsing errors end the consultation without an update and are logged.
- State updates: The evaluator does not receive the counselor’s system prompt, so style differences arise from the reply texts rather than explicit condition labels.This design is intended to blind the evaluator to the experimental condition.
- Outputs: Daily outputs include class means, non-attendance, isolation, closeness, agent initial and final states, and a complete event timeline.The timeline includes consultation text, replies, and evaluator explanations.
C Distributions of individual students on the first and final days
Figures 8–10 compare distributions of four main student parameters on the first and final days across basic, long-term, and high-frequency consultation settings. Students who stopped attending school are assigned happiness 0 on the final day.
- Figures 8, 9, and 10 show box plots of four main parameters across 20 students on each run’s initial and final days.The parameters are stress, happiness, self-reliance, and AI dependence.
- Students who stopped attending school appear with happiness 0 on the final day.
- The basic setting uses 15 days with a consultation threshold of 0.5, while the long-term setting extends to 50 days at the same threshold.
- The high-frequency setting uses 15 days and lowers the consultation threshold to 0.3.