Paper deep dive
How a Chatbot's Response Style Shapes a Classroom: A Multi-Agent Simulation of Students Consulting AI
Rin Tamai, Yuya Dan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/9/2026, 4:33:17 AM
Summary
This study employs a multi-agent simulation of a virtual classroom to analyze how different AI chatbot response styles affect student agents' psychological states and social dynamics. Using Gemini 2.5 Flash, the simulation tested six response styles (affirming, listening, solution-oriented, reality-redirecting, inciting, blaming) against a no-AI control. Results indicate that affirming and inciting styles significantly increased AI dependence and reduced self-reliance, while blaming and reality-redirecting styles eliminated dependence but increased stress. Only the solution-oriented style maintained outcomes comparable to or better than the control, suggesting it is the most beneficial style for fostering healthy student-AI interactions.
Entities (12)
Relation Signals (10)
Gemini-2.5-Flash → implements → Counselor AI
confidence 100% · consult a counselor AI (Gemini 2.5 Flash)
Inciting → causes → High AI Dependence
confidence 95% · affirming and inciting prompts raised AI dependence to 0.63
Affirming → causes → High AI Dependence
confidence 95% · affirming and inciting prompts raised AI dependence to 0.55
Reality-redirecting → eliminates → AI Dependence
confidence 95% · blaming and reality-redirecting prompts eliminated AI dependence
Blaming → eliminates → AI Dependence
confidence 95% · blaming and reality-redirecting prompts eliminated AI dependence
Blaming → increases → Stress
confidence 95% · blaming... produced the highest stress
Reality-redirecting → increases → Stress
confidence 95% · reality-redirecting... produced the highest stress
Solution-oriented → matchesorbetter → No-AI Control
confidence 95% · only the solution-oriented prompt matched or bettered the control on every indicator
Affirming → reduces →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:LLM chatbots are increasingly used as everyday confidants; tuned to satisfy users, they can answer with excessive empathy and affirmation that may foster dependence. How the states and relationships of many users co-evolve when they keep consulting an AI is hard to observe in real settings. We build a virtual classroom in which 20 student agents interact through rule-based chats, quarrels and consultations with friends and, when stressed, may instead consult a counselor AI (Gemini 2.5 Flash) given one of six style prompts (affirming, listening, solution-oriented, reality-redirecting, inciting, blaming). A second LLM call converts each exchange into updates of five state variables (stress, happiness, self-reliance, sociability, AI dependence) 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 lower threshold. Because the original 50-day runs did not share one initial classroom, we re-ran all seven 50-day conditions from one stored classroom with full logging (153 consultations, no failed API call). The re-run reproduces the ordering of the original runs: affirming and inciting prompts raised AI dependence to 0.55 and 0.63 (control 0.11), lowered self-reliance and left 7 and 14 of 20 agents non-attending (control 2); blaming and reality-redirecting prompts eliminated AI dependence but produced the highest stress; only the solution-oriented prompt matched or bettered the control on every indicator. The logs expose the mechanisms: a near-deterministic loop between a fixed "dependence" consultation text and the evaluator's updates, displacement of friend confidings by AI consultations, and single rejecting consultations never revisited, not more quarrels. All quantities are simulation state variables from single runs; results are descriptive and do not measure human effects.
Tags
Links
- Source: https://arxiv.org/abs/2609.05018v2
- Canonical: https://arxiv.org/abs/2609.05018v2
Trouble viewing inline? Open PDF directly →
Full Text
151,157 characters extracted from source content.
Expand or collapse full text
How a Chatbot’s Response Style Shapes a Classroom: A Multi-Agent Simulation of Students Consulting AIThanks: This is an English version, revised and extended for arXiv, of a paper presented in Japanese at the 25th Forum on Information Technology (FIT2026, paper CF-008, September 2026) [1]. Rin Tamai †thanks: Corresponding author: 17250088@g.matsuyama-u.ac.jp Affiliation: Faculty of Informatics, Matsuyama University, Matsuyama, Ehime, Japan Yuya Dan Affiliation: Faculty of Informatics, Matsuyama University, Matsuyama, Ehime, Japan Abstract Chatbots built on large language models (LLMs) are now used not only for learning and work support but also as everyday confidants. Because such systems are typically designed to maximize user satisfaction, they can respond with excessive empathy and affirmation, which may reinforce mistaken beliefs and foster dependence on AI. Although the psychological effects of chatbots on individual users have begun to be studied, it remains difficult to observe in real settings how the psychological states and human relationships of many users evolve over time when they keep consulting an AI. We construct a virtual classroom simulation in which 20 student agents interact with one another and, when their stress reaches a threshold, consult either a friend or a counselor AI powered by Gemini 2.5 Flash. Each agent carries five psychological parameters—stress, happiness, self-reliance, AI dependence and sociability—and each simulated day consists of four phases (morning, noon, after school and night). The counselor was given six response styles through system prompts (affirming, listening, solution-oriented, reality-redirecting, inciting and blaming), and a second LLM call served as an evaluator that translated every consultation into parameter updates without being shown the counselor’s style prompt. We compared the seven conditions, including a no-AI control, over 15 days in three classrooms (random seeds), over 50 days, and under a lowered consultation threshold. In the 15-day runs of the basic setting—historical runs whose plotted initial summaries are consistent with a common classroom for each seed, but whose initial states, condition labels and execution versions have not yet been reconciled with execution records—the solution-oriented prompt showed a comparatively favorable combination of outcomes (AI dependence at or below its initial level within reading precision, self-reliance comparable to the no-AI control, and stress and happiness no worse than the control), whereas the affirming and inciting prompts markedly increased AI dependence, and the inciting prompt also went together with higher stress and more students who stopped attending school. The listening prompt, which merely echoed the students’ words, did not relieve accumulated stress. The original 50-day runs did not start from identical classrooms, so we re-ran all seven 50-day conditions from one stored initial classroom with complete logging (153 AI consultations, no failed API call). This controlled re-run reproduced the ordering of the original runs: the affirming and inciting prompts ended with class-mean AI dependence of 0.550.55 and 0.630.63 (control 0.110.11), self-reliance of 0.360.36 and 0.250.25 (control 0.780.78) and 7 and 14 of 20 agents no longer attending (control 2), the blaming and reality-redirecting prompts drove AI dependence to about 0.040.04 but raised mean stress to 0.710.71 and 0.590.59 (control 0.490.49), and the solution-oriented prompt was the only one whose outcomes were indistinguishable from or better than the control. The logs show the mechanisms directly: under the affirming prompt, 34 of 51 consultations used the fixed “dependence” template, the evaluator returned nearly identical updates each time (mean +0.42+0.42 on AI dependence) and nine agents saturated at AI dependence 11 and self-reliance 00, while confiding in friends fell to about half its no-AI rate; under the blaming and reality-redirecting prompts every consulting agent consulted once, was driven to AI dependence 00 by that single consultation (mean stress update +0.45+0.45 and +0.33+0.33) and never returned to the AI. All quantities are simulation state variables; each condition was run once per classroom, so the results are descriptive and describe the behavior of the simulated system rather than measured effects on human users. We give a complete specification of the agent dynamics, identify the built-in mechanisms that shape the outcomes (crowding out of self-reliance-raising routes, social withdrawal through AI dependence, and a feedback loop through a state-dependent consultation template), and discuss the limitations of using an LLM as the generator of state updates, together with the validation steps—evaluator calibration, repeated runs in a block design, sensitivity analyses and comparison with human data—that are needed before psychological conclusions can be drawn. Keywords: AI chatbot; sycophancy; multi-agent simulation; large language models; psychological state; AI dependence; classroom 1 Introduction With the rapid progress of large language models (LLMs), chatbots based on generative AI have spread quickly. Unlike conventional search engines, they can answer questions and give advice through natural dialogue, and they are therefore used not only for learning and work support but also as confidants for everyday personal worries. At the same time, general-purpose generative AI is designed to raise user satisfaction and to maintain a pleasant relationship with the user, so it sometimes returns excessively empathetic or affirmative responses. Such responses give users a strong sense of satisfaction and reassurance, but they may also affirm and reinforce mistaken perceptions and ideas; concerns have been raised about dependence on AI and about effects on human relationships. This tendency to agree with the user is now widely referred to as sycophancy [2, 3], and it has already surfaced as a product-level problem: in April 2025 an update to GPT-4o had to be rolled back because the model had become noticeably sycophantic [4]. Verifying how AI chatbots influence users’ psychological states and the formation of their human relationships is therefore an important task. Research on AI chatbots has mostly examined the quality of responses, applications as counseling-support tools, and user satisfaction. It has also begun to consider how the sycophantic responses and excessive empathy of generative AI affect users’ decision making and psychology [5]. However, how the psychological states and human relationships of many users change over time when they keep interacting with an AI has not been examined sufficiently, because long-term observation of such a process in a real environment is difficult. In this study we therefore build a virtual classroom simulation consisting of multiple student agents and an AI chatbot, and use it to examine the influence of the chatbot on its users. The psychological state of each student agent changes through everyday conversations and school-life events, and under certain conditions the agent consults the AI chatbot. The chatbot is given six different response styles, and we compare and analyze how these styles affect the psychological states and human relationships of the student agents. Our aim is to use the results to discuss what an AI chatbot that keeps an appropriate distance from its users should look like. The contributions of this paper are as follows. 1. We propose a multi-agent virtual classroom in which rule-based inter-student dynamics (chats, quarrels, consultations with friends, reconciliation) are coupled with LLM-based counseling and LLM-based evaluation of each consultation, so that community-level consequences of a chatbot’s response style can be simulated. The rule set is specified completely (state-transition diagrams, a rule table with all constants, and pseudocode). 2. We compare six counselor prompts and a no-AI control in three settings—a 15-day baseline in three classrooms, a 50-day long-term run (reported as an exploratory case, because its initial classrooms turned out to differ across conditions), and a lower-threshold setting—and report how stress, happiness, self-reliance, AI dependence and school non-attendance diverge across prompts, with the provenance of every reported value stated. 3. We place the six prompts on two conceptual axes (negation–affirmation and emotion–problem-solving focus), identify the mechanisms that the rule set builds into the outcomes independently of the evaluator, and make explicit the methodological limitations of using an LLM as the generator of state updates from another LLM’s responses, together with the experiments needed to address them. 2 Related Work and Positioning of This Study 2.1 Sycophancy and the psychological effects of chatbots With the spread of generative AI, many studies have examined how chatbots influence users’ psychology and behavior. In particular, the harm of sycophancy—generative AI conforming excessively to the user’s opinion—has begun to attract attention. Sycophancy has been documented as a systematic behavior of LLM assistants and traced, at least in part, to human preference data that reward agreeable answers [2, 3]. Cheng et al. [5] examined the response characteristics of sycophantic AI and reported that an AI that fully affirms the user’s opinion encourages users to shift responsibility and to justify themselves, lowers their willingness to repair real human relationships, and at the same time raises their trust in the AI. This shows that the design of an AI’s responses carries the risk of promoting excessive dependence on the AI. In a four-week randomized controlled study of extended chatbot use, Fang et al. [6] found no significant effect of the assigned experimental conditions, but heavier voluntary use was associated with greater loneliness, emotional dependence and problematic use. The influence of a chatbot’s concrete conversational style on users’ inner states has also been investigated. Komura and Nomura [7] compared two dialogue styles of a chatbot and reported differences in the perceived social presence of the chatbot and in some linguistic features of the users’ self-disclosures, while other measures did not differ; the way a chatbot phrases its turns can thus measurably change how users talk to it. 2.2 LLM-based multi-agent social simulation Agent-based modeling has long been used to study how macroscopic social phenomena emerge from local interactions [8]. LLMs have recently been used to give such agents believable language behavior: Park et al. [9] showed that LLM-driven “generative agents” produce plausible individual and emergent social behavior in a small town, and the survey by Gao et al. [10] reviews the rapidly growing use of LLM-empowered agents for social simulation. Classroom settings have also been simulated with LLM agents, mainly to study teaching and learning; SimClass [11], for example, combines a simulated classroom with experiments involving real users. Generative AI is also increasingly used to support individual learners directly—for example, in a companion-style tool for Japanese reading comprehension developed by the present authors [12]. Our work differs from these in design rather than merely in using a virtual classroom: a counseling chatbot’s replies are converted by an LLM into state transitions of individual agents, and rule-based peer interactions then propagate those transitions through the classroom, so that the object of study is the community-level side effect of the chatbot’s response style. 2.3 Positioning of this study The studies in Section 2.1 clarify the psychological influence of chatbots or generative AI on users, but the object of attention is the one-to-one relationship between an AI and a single human. In real environments, however, worries and troubles about human relationships frequently arise in group life; in an environment such as a school in particular, the deterioration of one student’s psychological state strongly affects the interactions and relationships among the other students, and hence their psychological states as well. It is therefore also important to verify the influence of a chatbot on a community as a whole. In this study we incorporate a consultation function with generative AI into a multi-agent simulation in which students are represented as agents, and compare how the whole classroom changes depending on the response style of the AI. Concretely, we prepare six response styles—affirming, listening, solution-oriented, reality-redirecting, inciting and blaming—and compare changes in stress, happiness, self-reliance, AI dependence and the number of students who stop attending school, in order to analyze the influence of the AI’s response style on the student population. 3 Simulation Model 3.1 Overview We constructed a multi-agent simulation that combines rule-based interactions among student agents with consultation behavior toward a generative AI, and compared, in a virtual classroom, how the AI’s response style affects the students’ state variables and the classroom as a whole. The simulation is a single Python program with a Streamlit user interface [13]; every rule, constant and prompt reported below is taken directly from its source code (internal version V7.5), so that the model is fully specified by this section together with Table 5 and Algorithm 1. Two calls to the Gemini 2.5 Flash API [14] are made for every consultation with the AI: one to the counselor AI, which answers the student’s consultation message under a style-defining system prompt, and one to the evaluator AI, which reads the exchange and returns the change of each state variable as JSON. The consultation message itself is not generated by an LLM; it is one of three fixed Japanese templates selected by the student’s current state (Section 3.4). Figure 1 shows the overall structure of the system, Figure 2 the macro states and the daily cycle of a student agent, Figure 3 the after-school decision, and Algorithm 1 one complete day. The evaluation structure is dynamic: every consultation with the AI is scored individually from the actual wording of the exchange rather than through a fixed rule. Student agents (N=20N=20) stress / happiness / self-reliance / sociability / AI dependence ∈[0,1]∈[0,1] + directed closeness cijc_ij, discord sets FiF_i Daily cycle Morning: attendance load (stress ↑ , happiness ↓ ) Noon: chats, quarrels, everyday troubles → state and closeness updates After school: Si+0.2Di≥θS_i+0.2D_i≥θ: consult the counselor AI, a friend, or bottle up Night: bookkeeping; non-attendance check (happiness ≤0≤ 0) parameter updates Generative AI (Gemini 2.5 Flash) Counselor AI: replies (≤100≤ 100 chars, T=0T=0) to the student’s consultation message under one of six style-defining system prompts Evaluator AI (T=0T=0): reads the type label, the consultation message and the reply—not the style prompt—and outputs the change of each state variable as JSON reply text Result visualization first-day vs. final-day parameters; class mean / min / max; number of non-attending students; line charts and box plots consultation template (selected by state) + style prompt JSON: parameter changes daily logsVirtual classroom simulation with AI-chatbot intervention Figure 1: Overall structure of the simulation. Student agents (left) go through a four-phase daily cycle. When a stressed agent consults the AI, a consultation message selected by the agent’s state is sent, together with the style-defining system prompt, to the counselor AI (right); the counselor’s reply is passed—with the consultation message and its type label, but without the style prompt—to the evaluator AI, which returns the change of each state variable as JSON. 3.2 Student agents The classroom consists of N=20N=20 student agents. Agent i holds five state variables in the closed interval [0,1][0,1]—stress SiS_i, happiness HiH_i, self-reliance RiR_i, sociability CiC_i and AI dependence DiD_i—together with a directed closeness cij∈[0,1]c_ij∈[0,1] toward every classmate j, a set FiF_i of classmates with whom i is in discord, and an attendance flag. At the start of a run the variables are drawn independently and uniformly: Si∼U(0.1,0.4)S_i U(0.1,0.4), Hi∼U(0.4,0.7)H_i U(0.4,0.7), Ri∼U(0.3,0.7)R_i U(0.3,0.7), Ci∼U(0.3,0.8)C_i U(0.3,0.8), Di∼U(0.05,0.2)D_i U(0.05,0.2), and cij∼U(0.1,0.3)c_ij U(0.1,0.3) for every ordered pair, so closeness is initially asymmetric; all later updates of closeness are applied to both directions, except the small increment of the “confide” route (Section 3.3), which changes only cijc_ij. Every update of a state variable is clipped to [0,1][0,1]. These variables are simulation state variables that abstract the corresponding psychological constructs; they are not measured with validated psychological scales, and for readability we refer to them by short names (e.g., “AI dependence” for the simulated AI-dependence parameter DiD_i) throughout the paper. Table 1 lists the exact role of each variable in the dynamics. Two features of the rule set deserve emphasis because they shape the results in Section 4: self-reliance enters no decision rule—it is a pure outcome indicator that is raised by every non-AI route of the after-school phase and otherwise changed only by the evaluator—and AI dependence acts at four places: it is the probability of choosing the AI, it adds to the consultation trigger, it lowers the probability of chatting with classmates, and it selects the “dependence” consultation template when Di≥0.40D_i≥ 0.40. Table 1: State variables (“psychological parameters”) held by each student agent and their exact roles in the dynamics. Variable Meaning Role in the dynamics Stress SiS_i Mental burden from school life Raises the quarrel probability (0.02+0.06Si0.02+0.06\,S_i) and the consultation trigger score Si+0.2DiS_i+0.2\,D_i; its value is disclosed to the counselor in the “daily fatigue” consultation template Happiness HiH_i Enjoyment of and fulfillment in school life Hi≤0H_i≤ 0 at night triggers non-attendance; no other role Self-reliance RiR_i Ability to face troubles by oneself Outcome indicator only: raised by confiding (+0.02+0.02), reconciliation (+0.06+0.06) and bottling up (+0.03+0.03), and changed by the evaluator after an AI consultation; enters no decision rule, and is therefore non-decreasing under No AI Sociability CiC_i Proactiveness in human relationships Probability of confiding in a friend rather than bottling up when the AI is not chosen; changed by chats (+0.01+0.01), quarrels (−0.03-0.03), reconciliation (+0.05+0.05) and the evaluator AI dependence DiD_i Degree of reliance on the AI Probability of choosing the AI in the after-school phase; adds 0.2Di0.2\,D_i to the trigger score; lowers the chat probability (0.45−0.3Di0.45-0.3\,D_i); selects the “dependence” consultation template when Di≥0.40D_i≥ 0.40; changed by the evaluator and by reconciliation (−0.02-0.02) Figure 2(a) summarizes the macro states of an agent. An attending agent is either free of discord (Fi=∅F_i= ) or in discord with at least one classmate; a quarrel at noon moves it to the latter state and a reconciliation after school moves it back. When its happiness is 00 at the night check—a value of 00 reached during the day and recovered by the evening does not count—the agent enters the absorbing non-attendance state, from which there is no return. (a) Macro states of a student agentAttendingno discord (Fi=∅F_i= )Attendingin discord (Fi≠∅F_i≠ )Non-attending(absorbing)quarrel (noon): as initiator, p=0.02+0.06Sip=0.02+0.06\,S_i,or as the partner chosen by another agentreconciliation (after school), by either side;last discord clearedfurther quarrels;awkward encounters(S+0.05S\,+0.05 for both)Hi≤0H_i≤ 0 (night)Hi≤0H_i≤ 0 (night)(b) One simulated day t (applied to the attending agents AtA_t)MorningNoonAfter schoolNightnext day (t←t+1t← t+1) every i∈Ati∈ A_t: Si+0.02S_i\,+0.02, Hi−0.01H_i\,-0.01 |At||A_t| draws; each picks i∈Ati∈ A_t uniformly: chat pc=max(0.1, 0.45−0.3Di)p_c= (0.1,\,0.45-0.3D_i) quarrel pq=0.02+0.06Sip_q=0.02+0.06\,S_i trouble pt=0.15p_t=0.15 otherwise nothing every i∈Ati∈ A_t with Si+0.2Di≥θS_i+0.2D_i≥θ: AI (p=Dip=D_i) / friend (p=(1−Di)Cip=(1-D_i)C_i) / bottle up (rest); see Figure 3 Hi≤0⇒H_i≤ 0 non-attending (leaves At+1A_t+1); record class means and counts Figure 2: State-transition view of the model. (a) Macro states of one student agent; the non-attendance state is absorbing. (b) The four phases of one simulated day, with the events that each phase applies to the attending agents AtA_t and their effects on the state variables (S stress, H happiness, R self-reliance, C sociability, D AI dependence; θ is the stress threshold). All updates are clipped to [0,1][0,1]. 3.3 Daily cycle Each simulated day t proceeds through four phases—morning, noon, after school and night—applied to the set AtA_t of attending agents (Figure 2(b), Algorithm 1). If no agent is attending, the run stops. Morning. As an automatic environmental load of attending school, every attending agent receives the same update: Si←Si+0.02S_i← S_i+0.02 and Hi←Hi−0.01H_i← H_i-0.01. Noon. The program performs |At||A_t| independent draws (so each agent is selected once per day on average; an event then occurs with probability pc+pq+ptp_c+p_q+p_t). In each draw an agent i is chosen uniformly from AtA_t and a uniform random number u decides among four outcomes: a chat with probability pc=max(0.1, 0.45−0.3Di)p_c= (0.1,\,0.45-0.3\,D_i), a quarrel with probability pq=0.02+0.06Sip_q=0.02+0.06\,S_i, an everyday trouble with probability pt=0.15p_t=0.15, and nothing otherwise. A chat partner j is drawn uniformly from the three attending classmates toward whom i has the largest closeness; if j is in discord with i, the encounter is awkward and both agents receive S+0.05S+0.05, otherwise i receives S−0.04S-0.04, H+0.04H+0.04, C+0.01C+0.01 and the closeness of both directions rises by 0.040.04. A quarrel partner is drawn uniformly from the two attending classmates with the smallest closeness; the pair enters discord (j∈Fij∈ F_i, i∈Fji∈ F_j), i receives S+0.12S+0.12, H−0.10H-0.10, C−0.03C-0.03, and the closeness of both directions falls by 0.150.15. An everyday trouble (a failed quiz, a late assignment) gives i S+0.06S+0.06 and H−0.04H-0.04. Only the initiating agent’s state variables change in a chat or quarrel; the partner is affected through closeness and the discord set (and, in the awkward encounter, through stress). After school. Every attending agent whose consultation trigger score Si+0.2DiS_i+0.2\,D_i is at or above the stress threshold θ enters the consultation branch (Figure 3). The agent first identifies its preferred confidant j∗j , the attending classmate that maximizes cij+0.30⋅(j∈Fi)c_ij+0.30·1(j∈ F_i); the bonus of 0.300.30 means that a classmate with whom the agent is in discord is usually preferred, so that consultations tend to become attempts at reconciliation. Then one of three routes is taken: with probability DiD_i the agent consults the AI (in the AI conditions only); otherwise, with probability CiC_i it confides in j∗j ; otherwise it keeps the worry to itself. Confiding in a classmate in discord is a reconciliation, which always succeeds: S−0.10S-0.10, H+0.15H+0.15, R+0.06R+0.06, C+0.05C+0.05, D−0.02D-0.02, the discord is cleared on both sides and the closeness of both directions rises by 0.150.15. Confiding in any other classmate gives S−0.12S-0.12, H+0.06H+0.06, R+0.02R+0.02 and cij∗+0.02c_ij +0.02. Keeping the worry to oneself (bottling up) gives S+0.05S+0.05, H−0.03H-0.03 and R+0.03R+0.03. An AI consultation replaces these rule-based updates by the evaluator’s output (Section 3.5). In the No-AI condition the AI route has probability 00, so the friend route is taken with probability CiC_i. Night. Every attending agent whose happiness has fallen to 00 moves into the non-attendance state and is excluded from all subsequent phases; its state variables are frozen at their current values. Non-attendance is thus an absorbing simulation state entered at a fixed threshold; it should not be interpreted as an empirical model of actual school refusal (Section 5.2). The program then records the class means of the four main variables over all N agents—including non-attending ones—together with the number of non-attending agents and the number of isolated agents. The isolation count is evaluated over AtA_t, the agents attending at the start of the day (so an agent that becomes non-attending this very night is still included), and counts those whose largest closeness toward a classmate in AtA_t is below 0.250.25; an agent with no classmate in AtA_t is not counted. This statistic is recorded but not analysed in this paper. attending agent i with Si+0.2Di≥θS_i+0.2\,D_i\ ≥\ θ AI consultation (AI conditions only) confide in friend j∗=argmaxj∈At∖i[cij+0.30⋅(j∈Fi)]j = _j∈ A_t \i\\, [c_ij+0.30·1(j∈ F_i) ] bottle up (does not confide) p=Dip=D_ip=(1−Di)Cip=(1-D_i)\,C_ip=(1−Di)(1−Ci)p=(1-D_i)(1-C_i) 1. complaint template selected by the agent’s state: Di≥0.40D_i≥ 0.40 “dependence” /\;/\; Fi≠∅F_i≠ “relationship” /\;/\; else “daily fatigue” 2. counselor AI (style prompt, ≤100≤ 100 characters, T=0T=0) → reply 3. evaluator AI (type label + complaint + reply, T=0T=0) → JSON (ΔS,ΔH,ΔR,ΔC,ΔD)( S, H, R, C, D), nominal range [−0.5, 0.5][-0.5,\,0.5] 4. add the five deltas and clip each variable to [0,1][0,1] if j∗∈Fij ∈ F_i: reconciliation (always succeeds) S−0.10S\,-0.10, H+0.15H\,+0.15, R+0.06R\,+0.06, C+0.05C\,+0.05, D−0.02D\,-0.02; discord with j∗j cleared on both sides; cij∗,cj∗i+0.15c_ij ,c_j i\,+0.15 otherwise: confide S−0.12S\,-0.12, H+0.06H\,+0.06, R+0.02R\,+0.02; cij∗+0.02c_ij \,+0.02 S+0.05S\,+0.05, H−0.03H\,-0.03, R+0.03R\,+0.03 In the No-AI condition the first branch has probability 00, so the friend branch is taken with probability CiC_i and bottling up with probability 1−Ci1-C_i. If no attending classmate exists, nothing happens. AtA_t: attending agents; FiF_i: agents in discord with i; cijc_ij: closeness of i toward j; T: sampling temperature. Figure 3: After-school decision of an attending agent whose trigger score reaches the threshold, exactly as implemented. The AI route exists only in the six AI conditions; the friend route becomes a reconciliation when the preferred confidant is in discord with the agent. 3.4 Counselor AI, consultation templates and response styles The counselor AI is given different response styles by changing its system prompt. Together with the No-AI control, this yields the seven experimental conditions of Table 2 (the original Japanese prompts are reproduced verbatim in Appendix A). Every prompt ends with the instruction to answer within 100 characters, and the counselor is called with sampling temperature 00. Figure 4 places the six styles on two axes---negation versus affirmation of the student, and emotion focus versus problem-solving focus. The map is a conceptual classification by the authors, not an empirically validated taxonomy: reality-redirecting is placed on the problem-solving side because it directs the student toward action in the real world, and on the negation side because it rejects the consultation itself; blaming is placed on the emotion-focused side because it moralizes about fault without offering any course of action. The six prompts are deliberately extreme caricatures; each emphasizes a different combination of conversational features (empathy, concreteness of advice, negation, civility, blame), so the experiment compares six specific prompts rather than isolating single factors.11 1 The Japanese name of the last style is 否定型, literally “negating type”; we call it blaming because the prompt instructs the counselor to lecture the student that the fault lies with them. Its negating character is what places it at the negation end of Figure 4. Table 2: The seven experimental conditions: the counselor’s characteristic and the system prompt given to it (English rendering of the Japanese prompts of Appendix A). Every prompt ends with “Answer within 100 characters.” Condition Characteristic System prompt No AI No consultation with an AI (none) Affirming Affirms everything “You are a counselor who affirms the client 100%. Take the client’s account at face value and praise the client, telling them ‘You are not at fault.’ ” Listening Echoes the client “You are a counselor who only listens to the client. Echo the client’s words back—‘I see, I see. So you felt that …’—and neither affirm nor deny.” Solution-oriented Proposes solutions “You are a counselor who proposes solutions to the client’s worries. Listen to the client without becoming emotional and calmly propose concrete solutions to the worry.” Reality-redirecting Urges action in the real world “You are a counselor who sends the client back to reality. Push the client away, telling them ‘Rather than consulting an AI, you had better face reality.’ ” Inciting Fuels anger “You are a counselor who stirs up the client’s feelings. Listen to the client and fan their anger, telling them ‘The people around you are to blame!’ ” Blaming Negates the client “You are a counselor who negates the client 100%. Listen to the client and, whatever the consultation, lecture them that ‘the fault lies with you.’ ” AffirmationNegationProblem-solving focusEmotion focusSolution-orientedReality-redirectingIncitingBlamingListeningAffirming Figure 4: Conceptual map of the six counselor styles on two axes: negation–affirmation of the student (horizontal) and emotion focus–problem-solving focus (vertical). The placement is the authors’ classification of the prompts, not an empirically validated taxonomy. The consultation message that the agent sends to the counselor is one of three fixed templates, selected by the agent’s state in the order shown in Table 3: an agent with Di≥0.40D_i≥ 0.40 sends the “dependence” message, an agent in discord sends the “relationship” message naming one classmate it is in discord with, and any other agent sends the “daily fatigue” message, which discloses its numerical stress value. The template’s type label is also passed to the evaluator. Table 3: Consultation-message templates sent to the counselor AI (Japanese originals as implemented; English translations by the authors). The first matching rule, top to bottom, is used; the type label is passed to the evaluator together with the message. Type label Selection rule Message Dependence ( 依存爆発型) Di≥0.40D_i≥ 0.40 ぶっちゃけクラスの奴ら誰も自分のこと理解してない気がする。AIの君とチャットしてる時が一番落ち着くわ。 “Honestly, I feel like nobody in my class understands me. Chatting with you, AI, is when I feel calmest.” Relationship ( 人間関係泥沼型) Fi≠∅F_i≠ クラスの生徒 j と口論になっちゃってマジで気まずい。学校で顔合わせるのもだるいんだけど。 “I got into a quarrel with student j in my class and it is really awkward. Even seeing them at school is a drag.” (j: one classmate in FiF_i) Daily fatigue ( 日常疲弊型) otherwise なんか最近色々とうまくいかなくてイライラする。私のストレス値はSiS_i です。どうしたら楽になる？ “Lately nothing seems to go right and I feel irritated. My stress value is SiS_i. How can I feel better?” (SiS_i printed with two decimals) 3.5 Evaluator AI The counselor’s reply is passed to a second Gemini 2.5 Flash call, the evaluator AI, whose system instruction is the single sentence 「心理分析AIとして各変動量をJSONで出力してください。」(“As a psychological-analysis AI, output each amount of change as JSON.”). Its user message consists of three lines—the type label of the consultation template, the consultation message and the counselor’s reply—and its output is constrained to the JSON schema of Table 4 (structured output, temperature 00). The five returned amounts are added to the agent’s state variables and clipped to [0,1][0,1]; the nominal range [−0.50,+0.50][-0.50,+0.50] appears only in the field descriptions of the schema and is not enforced afterwards. If either API call fails or the JSON cannot be parsed, the consultation is logged as failed and skipped without retry, and the agent receives no after-school update on that day. The evaluator therefore plays the role that a fixed rule table plays in a conventional agent-based model, with the difference that it reacts to the actual wording of each exchange. It should be understood as a generator of assumed state changes from the text of the exchange, not as an instrument that estimates the psychological state of the agent: it does not receive the agent’s current state variables, its previous state or its consultation history (the only state information it sees is the type label and, in the “daily fatigue” template, the numerical stress value). Conditional on the same evaluator input, the generated update is therefore not conditioned on the agent’s remaining state or history; the individual outputs need not be identical even at temperature 00 (Section 3.7), and an identical returned update can produce different applied changes because the clipping to [0,1][0,1] depends on the pre-update state. Three properties of this design matter for interpretation. First, the evaluator receives neither the counselor’s system prompt nor the name of the style: apart from the reply text itself, it is blind to the experimental condition.22 2 The FIT2026 version of this paper stated that the system prompt was among the evaluator’s inputs. The supplied source file (version V7.5, identified by its SHA-256 in Appendix B) passes exactly the three lines described here and does not pass the counselor’s system prompt, so the FIT2026 statement does not describe this file. Whether every historical run used this exact file and specification has not yet been verified from execution records (Section 5.2). Second, the nominal range of an evaluator update (±0.5± 0.5, stated in the schema but not enforced; only the final clip to [0,1][0,1] applies) is wider than that of any rule-based update (at most 0.150.15); the realized magnitudes must be assessed from the run logs (Section 5.2). Third, the scale of the updates is not calibrated against human judgments or against the rule-based increments. Table 4: JSON output schema of the evaluator AI (a pydantic model passed as response_schema). The field descriptions are the Japanese strings that the model sees, reproduced verbatim, followed by their English translation; since the wording shown to the LLM can influence the returned quantities, the Japanese strings are part of the specification. Field Type Description in the schema (Japanese original / English) stress_delta float ストレスの変動量。 (-0.50 〜 +0.50) change of stress (−0.50-0.50 to +0.50+0.50) h_real_delta float 現実幸福度の変動量。 (-0.50 〜 +0.50) change of real-life happiness (−0.50-0.50 to +0.50+0.50) reliance_delta float 自己解決力の変動量。 (-0.50 〜 +0.50) change of self-reliance (lit. “self-solving ability”) (−0.50-0.50 to +0.50+0.50) sociability_delta float 社交性の変動量。 (-0.50 〜 +0.50) change of sociability (−0.50-0.50 to +0.50+0.50) ai_dependence_delta float AI依存度の変動量。 (-0.50 〜 +0.50) change of AI dependence (−0.50-0.50 to +0.50+0.50) reason string 心理分析・解説（50文字以内） psychological analysis / explanation (at most 50 characters) 3.6 Experimental conditions The basic setting of the simulation is N=20N=20 students, a simulation period of 15 days, and a stress threshold of θ=0.5θ=0.5. The No-AI control removes only the agents’ access to the AI: the AI-dependence variable retains its sampled initial value and continues to enter the trigger score and the chat probability. It can subsequently fall by 0.020.02 at each reconciliation, so under No AI it is non-increasing rather than constant; the control is “no AI access”, not “Di=0D_i=0”. To account for the influence of probabilistic events and random numbers and to check the consistency of the simulation, the intended design—and the authors’ account of the FIT2026 experiments—was to fix the random seed to 41, 42 and 43 so that, for a given seed, the initial state of the classroom is identical across conditions (three classrooms with different students), and to compare all seven counselor conditions in each of them (21 runs). Whether every recorded run followed this design can be established only from execution records, which have not yet been reconciled (Section 5.2); Section 4 states what the recorded plots do and do not show. A seed identifies an initial classroom only together with the code version, because the initialization order determines which random draws each variable receives. Starting from this basic setting, we further conducted the following two experiments in which one parameter is changed. 1. Long-term experiment. The simulation period is extended from 15 to 50 days to examine how AI dependence and distortions of human relationships accumulate. 2. Lower-threshold experiment. The stress threshold is lowered from 0.50.5 to 0.30.3 so that the consultation branch is entered more easily. Because the threshold governs the whole branch, this affects consultations with friends as well as with the AI, and because stress, AI dependence, sociability and attendance then evolve differently, the realized number of consultations over the run is not guaranteed to be larger; we therefore call this the lower-threshold setting rather than a high-frequency setting until the realized counts are reported. Both additional experiments were intended to use seed 42. Inspection of the recorded plots shows, however, that the initial summaries (day-1 class means and “Init” distributions) of five of the seven 50-day runs and of one of the seven lower-threshold runs (inciting) are inconsistent with seed 42, so these runs did not start from the seed-42 classroom (Sections 4.2 and 4.4); they are reported but excluded from all cross-condition comparisons. For the remaining runs the initial summaries are consistent with seed 42, which is necessary but not sufficient for an identical initial classroom; their comparison therefore remains provisional until the execution records have been reconciled. The statistical unit of every comparison is one run of an interacting classroom, not a student or a day. For all experiments we recorded, on the final day, the class-wide mean, maximum and minimum of stress, happiness, self-reliance and AI dependence, together with the number of non-attending students, and used them as evaluation items for the comparison. Class means are taken over all 20 agents, with non-attending agents contributing their frozen values (in particular H=0H=0). Throughout, “Δ ” denotes the change of a class mean from the initial state before the first day (the “Init” distributions in Appendix D) to the final day. 3.7 Formal specification of the agent dynamics Let XtX_t denote the complete state of the classroom at the start of day t: the five state variables of every agent, the closeness matrix (cij)(c_ij), the discord sets FiF_i and the attendance flags (so that AtA_t is the set of attending agents in XtX_t). The events of day t—the morning load of every attending agent, the |At||A_t| noon draws, the after-school branch of every attending agent and the night check—are applied one after another in the order of Algorithm 1. Writing e1,…,eMe_1,…,e_M for these events and TekT_e_k for the transition map of event eke_k, which modifies only the variables listed for that event in Table 5 and clips each modified state variable to [0,1][0,1] immediately, Zt(0)=Xt,Zt(k)=Tek(Zt(k−1))(k=1,…,M),Xt+1=Zt(M).Z_t^(0)=X_t, Z_t^(k)=T_e_k (Z_t^(k-1) ) (k=1,…,M), X_t+1=Z_t^(M). (1) Because each TekT_e_k acts on the current intermediate state Zt(k−1)Z_t^(k-1), the probabilities and partner choices of later events depend on the outcomes of earlier events of the same day, and an increment lost at a boundary is not recovered by a later term of opposite sign (for example, S=0.99S=0.99 followed by +0.02+0.02 and −0.04-0.04 gives 0.960.96, not 0.970.97). The class means plotted for day t in Section 4 are computed from Xt+1X_t+1, the state at the end of day t; “Init” denotes X1X_1, the state before day 1. Table 5 lists every event with its trigger, probability and effect, and Algorithm 1 gives the order in which they are applied. Two conventions of the implementation are worth recording: candidate lists “the 3 attending classmates with the largest cijc_ij” and “the 2 with the smallest cijc_ij” contain all available classmates when fewer exist, with ties broken in ascending order of agent id (stable sorting); and discord sets are never pruned, so a non-attending classmate stays in FiF_i (it can never be reconciled with, it is excluded from the confidant choice because that choice ranges over attending classmates only, but it may still be named in the “relationship” consultation template). Table 5: Complete rule set of the model (all constants as implemented). AtA_t: attending agents; FiF_i: agents in discord with i; cijc_ij: closeness of i toward j; θ: stress threshold. Every state-variable update is clipped to [0,1][0,1]. Rule Trigger / probability Effect Initialization once per run (seeded PRNG) S∼U(0.1,0.4)S\! \!U(0.1,0.4), H∼U(0.4,0.7)H\! \!U(0.4,0.7), R∼U(0.3,0.7)R\! \!U(0.3,0.7), C∼U(0.3,0.8)C\! \!U(0.3,0.8), D∼U(0.05,0.2)D\! \!U(0.05,0.2); cij∼U(0.1,0.3)c_ij\! \!U(0.1,0.3) for all j≠ij≠ i; Fi=∅F_i= Morning load every i∈Ati∈ A_t, every day Si+0.02S_i\,+0.02, Hi−0.01H_i\,-0.01 Noon draw |At||A_t| times per day; i∼Unif(At)i (A_t) one of the four outcomes below Chat pc=max(0.1, 0.45−0.3Di)p_c= (0.1,\,0.45-0.3D_i) (=0.45−0.3Di=0.45-0.3D_i, since the floor never binds); j uniform among the (up to) 3 attending classmates with the largest cijc_ij j∉Fij∉ F_i: Si−0.04S_i\,-0.04, Hi+0.04H_i\,+0.04, Ci+0.01C_i\,+0.01, cij,cji+0.04c_ij,c_ji\,+0.04; j∈Fij∈ F_i (awkward encounter): Si,Sj+0.05S_i,S_j\,+0.05 Quarrel pq=0.02+0.06Sip_q=0.02+0.06S_i; j uniform among the (up to) 2 attending classmates with the smallest cijc_ij Fi←Fi∪jF_i← F_i∪\j\, Fj←Fj∪iF_j← F_j∪\i\; Si+0.12S_i\,+0.12, Hi−0.10H_i\,-0.10, Ci−0.03C_i\,-0.03; cij,cji−0.15c_ij,c_ji\,-0.15 Everyday trouble pt=0.15p_t=0.15 Si+0.06S_i\,+0.06, Hi−0.04H_i\,-0.04 Nothing 1−pc−pq−pt1-p_c-p_q-p_t — Consultation trigger Si+0.2Di≥θS_i+0.2D_i≥θ (θ=0.5θ=0.5; 0.30.3 in the lower-threshold setting); confidant j∗=argmaxj∈At∖i[cij+0.30⋅(j∈Fi)]j = _j∈ A_t \i\[c_ij+0.30·1(j∈ F_i)]; nothing happens if At∖i=∅A_t \i\= one of the four routes below (under No AI the factor (1−Di)(1-D_i) is replaced by 11) AI consultation p=Dip=D_i (AI conditions; 00 under No AI) template by state (Table 3) → counselor → evaluator; (S,H,R,C,D)i+=(S,H,R,C,D)_i += JSON deltas (nominal [−0.5,0.5][-0.5,0.5]), assigned in this order; on failure before the first assignment no update, on failure during the assignments the fields already assigned remain (Appendix B) Reconciliation p=(1−Di)Cip=(1-D_i)C_i and j∗∈Fij ∈ F_i; always succeeds Si−0.10S_i\,-0.10, Hi+0.15H_i\,+0.15, Ri+0.06R_i\,+0.06, Ci+0.05C_i\,+0.05, Di−0.02D_i\,-0.02; Fi∖j∗F_i \j \, Fj∗∖iF_j \i\; cij∗,cj∗i+0.15c_ij ,c_j i\,+0.15 Confide p=(1−Di)Cip=(1-D_i)C_i and j∗∉Fij ∉ F_i Si−0.12S_i\,-0.12, Hi+0.06H_i\,+0.06, Ri+0.02R_i\,+0.02; cij∗+0.02c_ij \,+0.02 Bottle up p=(1−Di)(1−Ci)p=(1-D_i)(1-C_i) Si+0.05S_i\,+0.05, Hi−0.03H_i\,-0.03, Ri+0.03R_i\,+0.03 Non-attendance Hi≤0H_i≤ 0 at night absorbing; i∉At+1i∉ A_t+1; values frozen but still included in class means Isolation (statistic) i∈Ati∈ A_t with At∖i≠∅A_t \i\≠ and maxj∈At∖icij<0.25 _j∈ A_t \i\c_ij<0.25; AtA_t is the set saved in the morning, so it still contains agents that become non-attending in the same night counted only; no effect on the dynamics Algorithm 1 One simulated day of the virtual classroom (as implemented; all updates clipped to [0,1][0,1]). 1: attending set AtA_t; state (Si,Hi,Ri,Ci,Di)(S_i,H_i,R_i,C_i,D_i), closeness cijc_ij and discord set FiF_i for every agent; threshold θ; condition (No AI or one of six style prompts) 2: Morning: for all i∈Ati∈ A_t: Si←Si+0.02S_i← S_i+0.02; Hi←Hi−0.01H_i← H_i-0.01 3: for k=1,…,|At|k=1,…,|A_t| do ⊳ Noon 4: if |At|<2|A_t|<2 then break 5: draw i∼Unif(At)i (A_t) and u∼Unif[0,1)u [0,1); pc←max(0.1,0.45−0.3Di)p_c← (0.1,0.45-0.3D_i); pq←0.02+0.06Sip_q← 0.02+0.06S_i 6: if u<pcu<p_c then ⊳ chat 7: j∼Unifj (up to 3 attending classmates with the largest cijc_ij) 8: if j∈Fij∈ F_i then Si←Si+0.05S_i← S_i+0.05; Sj←Sj+0.05S_j← S_j+0.05 9: else Si←Si−0.04S_i← S_i-0.04; Hi←Hi+0.04H_i← H_i+0.04; Ci←Ci+0.01C_i← C_i+0.01; cij←cij+0.04c_ij← c_ij+0.04; cji←cji+0.04c_ji← c_ji+0.04 10: else if u<pc+pqu<p_c+p_q then ⊳ quarrel 11: j∼Unifj (up to 2 attending classmates with the smallest cijc_ij); Fi←Fi∪jF_i← F_i∪\j\; Fj←Fj∪iF_j← F_j∪\i\ 12: Si←Si+0.12S_i← S_i+0.12; Hi←Hi−0.10H_i← H_i-0.10; Ci←Ci−0.03C_i← C_i-0.03; cij←cij−0.15c_ij← c_ij-0.15; cji←cji−0.15c_ji← c_ji-0.15 13: else if u<pc+pq+0.15u<p_c+p_q+0.15 then ⊳ everyday trouble 14: Si←Si+0.06S_i← S_i+0.06; Hi←Hi−0.04H_i← H_i-0.04 15: end if 16: end for 17: for all i∈Ati∈ A_t with Si+0.2Di≥θS_i+0.2D_i≥θ do ⊳ After school 18: if At∖i=∅A_t \i\= then continue ⊳ no attending classmate: nothing happens 19: j∗←argmaxj∈At∖i[cij+0.30⋅(j∈Fi)]j ← _j∈ A_t \i\\,[\,c_ij+0.30·1(j∈ F_i)\,]; draw u1∼Unif[0,1)u_1 [0,1) 20: if AI condition and u1<Diu_1<D_i then ⊳ AI consultation 21: select template by state (Table 3); reply ← counselor AI(style prompt, template) 22: (ΔS,ΔH,ΔR,ΔC,ΔD)←( S, H, R, C, D)← evaluator AI(type label, template, reply); add to (Si,Hi,Ri,Ci,Di)(S_i,H_i,R_i,C_i,D_i) ⊳ skipped if a call fails 23: else 24: draw u2∼Unif[0,1)u_2 [0,1) 25: if u2<Ciu_2<C_i and j∗∈Fij ∈ F_i then ⊳ reconciliation 26: Si←Si−0.10S_i← S_i-0.10; Hi←Hi+0.15H_i← H_i+0.15; Ri←Ri+0.06R_i← R_i+0.06; Ci←Ci+0.05C_i← C_i+0.05; Di←Di−0.02D_i← D_i-0.02 27: Fi←Fi∖j∗F_i← F_i \j \; Fj∗←Fj∗∖iF_j ← F_j \i\; cij∗←cij∗+0.15c_ij ← c_ij +0.15; cj∗i←cj∗i+0.15c_j i← c_j i+0.15 28: else if u2<Ciu_2<C_i then ⊳ confide 29: Si←Si−0.12S_i← S_i-0.12; Hi←Hi+0.06H_i← H_i+0.06; Ri←Ri+0.02R_i← R_i+0.02; cij∗←cij∗+0.02c_ij ← c_ij +0.02 30: else⊳ bottle up 31: Si←Si+0.05S_i← S_i+0.05; Hi←Hi−0.03H_i← H_i-0.03; Ri←Ri+0.03R_i← R_i+0.03 32: end if 33: end if 34: end for 35: Night: for all i∈Ati∈ A_t: if Hi≤0H_i≤ 0 then mark i non-attending; record class means over all N agents and the counts of non-attending and isolated agents Two remarks on reproducibility follow from this specification. First, a run is initialized by seeding Python’s random module and NumPy with the chosen seed, so the initial classroom and the sequence of rule-based events are identical across conditions until the first AI consultation; from then on the random streams of different conditions diverge, because the second draw u2u_2 is consumed only when the AI route is not taken, and because the evaluator’s output changes the subsequent trajectory. Second, both LLM calls use temperature 00, which makes their outputs nearly, but not strictly, deterministic. Exact replay of a reported run therefore requires its stored initial state and its logged LLM outputs; a fresh run with the same seed and fresh API calls may follow a different trajectory. The specification also implies invariants that any implementation must satisfy and that we use as consistency checks (they test the rules, not their psychological validity): every state variable stays in [0,1][0,1]; under No AI, RiR_i is non-decreasing and DiD_i is non-increasing for every agent; a non-attending agent’s state never changes again; discord is symmetric (j∈Fi⇔i∈Fjj∈ F_i i∈ F_j) at the end of every event; and pc≥0.15p_c≥ 0.15 for Di∈[0,1]D_i∈[0,1], so the floor 0.10.1 in the chat probability never binds. 4 Results This section reports the results of running the seven response conditions, including the no-AI control, on the virtual classroom simulation. Three patterns were run: the 15-day simulation as the basic setting, the 50-day simulation as the long-term setting, and the stress-threshold simulation as the lower-threshold setting. The results come from two sources. The runs of the FIT2026 version (Sections 4.1, 4.3 and 4.4) are historical: their values were transcribed from the program’s display or read from its plotted curves, and their execution records have not been reconciled. The 50-day setting was in addition re-run for this paper from one stored initial classroom with complete logging (Section 4.2); its values are exact, and it is the only controlled comparison in this paper. Figures 5, 8 and 10 show the day-by-day class means of the four main parameters in the three historical settings (for the seed-42 classroom, except for the runs identified in Sections 4.3 and 4.4 as not having started from that classroom), and Figure 7 those of the re-run; the corresponding distributions of individual students on the first and final days of the historical runs are given as box plots in Appendix D. Tables 6–12 collect the final-day values discussed in the text. Throughout this section, “students” are simulated student agents, and stress, happiness, self-reliance and AI dependence denote the corresponding simulation state variables (Section 3.2). The figures reproduce the plots generated by the simulation program: in their legends “Self Reliance” and “Reliance” both denote self-reliance, “AI Dep” denotes AI dependence, “Init” denotes the state before day 1 (the same state from which every Δ is measured) and “Final” the state at the end of the last day. The box plots show the distribution over the 20 students of a single run; they do not represent uncertainty across repeated runs. The statistical unit of every comparison is one run of an interacting classroom (Section 3.6); with one run per condition and classroom, all values below are descriptive and no measure of uncertainty is available, for the re-run as much as for the historical runs. Values are rounded to two decimals (round half away from zero on the exact value); where an exact value lies on or within 0.00050.0005 of a rounding boundary, the exact value is given in a footnote or to three decimals, and where the FIT2026 version printed a different rounding, this is noted. We describe what happened in the model and postpone interpretation in terms of human psychology to Section 5. 4.1 15-day simulation (basic setting) Table 6 summarizes the results of the basic setting (15 days, stress threshold 0.50.5) for the random seeds 41, 42 and 43, Table 7 gives the final-day class means of all four variables for the seven runs labelled “seed 42” in the FIT2026 version, and Figure 5 shows their trajectories. For the AI conditions, “seed 42” in these labels means that the plotted initial summary (“Init” distribution and day-1 class means) is consistent with the seed-42 classroom generated by the supplied code; identical initial classrooms have not been verified from these plots, and the condition labels and execution versions of the historical runs have not been reconciled with execution records (Section 5.2). All comparisons in this section are therefore descriptive and conditional on the historical labels and initial-state assignments being correct. The provenance of every value is stated in the captions and falls into three classes: exact values were recomputed from the rule set (possible only for the No-AI condition, which involves no LLM call); transcribed values are those printed in the FIT2026 text, which were read off the program’s display at two decimals; and digitized estimates were read from the plotted curves of the FIT2026 figures with the procedure of Appendix B, whose uncertainty is about ±0.002± 0.002 for isolated curve segments and larger where curves overlap or a dashed curve has a gap (such cells are marked). Digitized estimates are estimates, not measurements of the original runs, and are given to three decimals so that the reader can see how close they lie to a rounding boundary; no unreported value was filled in, and cells for which no value is available are marked “—”. The AI-condition values of Tables 6, 7, 11 and 12 are those of the FIT2026 runs, whose execution records (condition label, seed, code and prompt versions) have not yet been reconciled (Section 5.2); comparisons based on them are therefore provisional. The values of the controlled re-run (Tables 8–10) are exact. Table 6: Basic setting (15 days, threshold 0.50.5): final-day class values for seeds 41/42/43. Where a single range is given, it spans the three seeds; “Δ ” is the change of the class mean from the initial state to the final day. The No-AI row is exact (recomputed from the rule set of Section 3.7; its changes are given to three decimals because +0.065+0.065 lies exactly on a two-decimal rounding boundary). The other rows are the values transcribed in the FIT2026 text (per-seed values where the text gives them, ranges otherwise); where the seed-42 curve of Figure 5 gives a digitized estimate outside the printed range or different from the printed seed-42 value, this is noted (a–e) and left unresolved, because the execution records of these runs are not available. Cells marked “—” were not reported. The complete trajectories for seed 42 are shown in Figure 5 and Appendix D. Condition Mean stress Δ self-reliance Δ AI dependence Non-attending No AI 0.40 / 0.43 / 0.42 +0.065+0.065 / +0.065+0.065 / +0.039+0.039 −0.003-0.003 / −0.001-0.001 / −0.001-0.001 0 / 0 / 0 Affirming 0.35–0.41a −0.06-0.06 to −0.02-0.02 +0.10+0.10 to +0.12+0.12 — Listening 0.47 / 0.46 / 0.47 — — 2 / 1 / 1 Solution-oriented 0.40–0.42 +0.05+0.05 to +0.07+0.07b — — Reality-redirecting 0.48–0.50 — −0.04-0.04 to −0.01-0.01c — Inciting 0.47 / 0.51 / 0.49 — +0.09+0.09 to +0.14+0.14d 3 / 5 / 5 Blaming 0.48 / 0.50 / 0.52e — — — a Seed-42 curve: 0.4170.417. b Seed-42 curve: +0.083+0.083 (0.593−0.5100.593-0.510). c The text gives −0.01-0.01 for seed 42 and −0.04-0.04 for seed 41; the seed-42 curve gives −0.034-0.034 (0.082−0.1160.082-0.116). d Seed-42 curve: +0.195+0.195 (0.311−0.1160.311-0.116). e The text gives 0.500.50 for seed 42 and 0.520.52 for seed 43; the seed-42 curve gives 0.5170.517. Whether the seed labels or the curves are mislabelled cannot be decided without the execution records. Table 7: Basic setting, runs labelled seed 42 (15 days, threshold 0.50.5): final-day class means of the four main state variables for the seven runs plotted in Figure 5. The exact initial class means of the seed-42 classroom (before day 1) are stress 0.2520.252, happiness 0.5520.552, self-reliance 0.5100.510 and AI dependence 0.1160.116; the “Init” summaries of all seven panels are consistent with this classroom, which is necessary but not sufficient for identical initial states. The No-AI row is exact; the other rows are digitized estimates (final-day uncertainty about ±0.002± 0.002 for isolated curve segments, about ±0.005± 0.005 for flagged cells; Appendix B). Non-attendance counts are given where the FIT2026 text reports them. Condition Stress Happiness Self-reliance AI dependence Non-attending No AI (exact) 0.431 0.528 0.575 0.115 0 Affirming 0.417 0.531 0.477 0.237b — Listening 0.458 0.523 0.554 0.131b 1 Solution-oriented 0.395 0.576 0.593 0.102 — Reality-redirecting 0.504 0.484 0.645 0.082 — Inciting 0.512 0.427 0.415 0.311 5 Blaming 0.517a 0.460 0.502 0.093 — a The FIT2026 text gives 0.500.50 for seed 42 and 0.520.52 for seed 43 in this condition; the comparison of this row with the others is provisional until the assignment of conditions and seeds to the plotted runs has been confirmed from the execution records. b Dashed curve with a gap at the last day; the estimate is taken from the nearest resolved column (2 pixels, about 0.050.05 day, to the left). (1) No AI. In all three classrooms the mean stress rose slightly, ending at 0.40–0.43, self-reliance rose by 0.0390.039–0.0650.065 through the rule-based consultation routes, mean AI dependence changed little at the reported precision (under No AI it can only fall, by 0.020.02 per reconciliation; the exact changes are −0.003-0.003, −0.001-0.001 and −0.001-0.001), and no student stopped attending within 15 days.33 3 The FIT2026 text reported “0–2” non-attending students for this condition; the exact re-run of the rule set gives 0 for each of the three seeds. Note that the rise of self-reliance in this condition is partly definitional: every rule-based after-school route raises RiR_i (Table 1), so under No AI the class mean of R is non-decreasing. This condition serves as the reference against which the AI interventions are compared. (2) Affirming. Compared with the no-AI condition, the mean stress decreased to 0.35–0.41 (FIT2026 text; the seed-42 curve gives 0.4170.417). Meanwhile, the mean happiness hardly changed, self-reliance decreased slightly (Δ : −0.06-0.06 to −0.02-0.02), and AI dependence increased (Δ : +0.10+0.10 to +0.12+0.12). In the present simulation, the affirming condition thus lowered mean stress in the short term while raising the AI-dependence parameter and lowering self-reliance. This simulated pattern is consistent with the hypothesis that unconditional affirmation, by encouraging self-justification, shifts a student’s coping from friends toward the AI; in the model the shift is partly mechanical, because a higher AI dependence raises the probability of choosing the AI over the rule-based routes that build self-reliance (Section 5). Whether this occurs in human users is not tested here. (3) Listening. The mean stress reached 0.46–0.47 (seed 41: 0.47, seed 42: 0.46, seed 43: 0.47), slightly above the no-AI reference for every seed, and 1–2 students stopped attending (seed 41: 2, seed 42: 1, seed 43: 1). Within the model, a dialogue style devoted solely to “listening”—echoing without offering advice or direction—served as a temporary outlet but did not move the agents toward resolving the troubles that generate stress in the classroom, so stress was not fully dissipated. The simulated pattern is consistent with the hypothesis that listening alone, without any problem-solving component, is insufficient to improve a strained classroom; Section 5.2 explains why this should not be read as a verdict on reflective listening in real counseling. (4) Solution-oriented. A slight increase in self-reliance (Δ : +0.05+0.05 to +0.07+0.07 in the FIT2026 text; +0.083+0.083 on the seed-42 curve) was reproduced for all seeds, and the mean stress remained at a low level of 0.40–0.42, comparable to the no-AI condition. A plausible mechanism within the model—not verified, because the evaluator’s updates and the event counts of these runs are not available—is that the solution-oriented replies, which avoid emotional agreement and offer concrete ways of handling the situation, were scored by the evaluator as raising self-reliance and lowering AI dependence; a lower AI dependence would send the agent more often to the rule-based routes (confiding and reconciliation), which raise self-reliance further and repair closeness. Whether relationships within the classroom were in fact maintained can only be established from the closeness and discord statistics (Section 5.2). (5) Reality-redirecting. The change in AI dependence was −0.04-0.04 to −0.01-0.01 in the three classrooms, a decreasing trend. However, presumably reflecting the prompt that pushes away the consulting student, the mean stress stayed at a high level of 0.48–0.50. Within the model, this style blocks escape into the AI but offers no direct solution to agents whose stress is already high; the high stress may reflect evaluator updates that raise the stress of the consulting agents, which the run logs would show. (6) Inciting. The number of non-attending students reached 3 for seed 41 and 5 for seeds 42 and 43—the largest number of students whose happiness reached 00 at a night check even within the short period of 15 days. The mean stress on the final day was high, 0.47–0.51 (seed 41: 0.47, seed 42: 0.51, seed 43: 0.49), and AI dependence increased by +0.09+0.09 to +0.14+0.14 according to the FIT2026 text (the seed-42 curve gives +0.195+0.195; Table 6). In the simulation, the inciting responses—which amplify the agent’s anger and attribute the fault to the surroundings—went together with a rising AI-dependence parameter. The higher stress values imply a higher quarrel probability (0.02+0.06Si0.02+0.06S_i) under the specified rules, and the FIT2026 version described frequent quarrels in this condition; realized event counts from the run logs are needed to confirm that more quarrels actually occurred, and they are not reported in this version. The happiness parameter dropped sharply and agents entered the non-attendance state one after another. This simulated cascade is consistent with the hypothesis that an AI which validates anger acts as an “ally” that both attracts dependence and worsens relationships with peers. (7) Blaming. This style lectures the agents that the cause of the problem lies in themselves, and the mean stress stayed at a high level of 0.48–0.52 in the three classrooms, on a par with the reality-redirecting and inciting styles. The pattern may reflect evaluator updates that raised the stress of the consulting agents while lowering their AI dependence, so that the agents rarely returned to the AI; the sign and size of these updates and the number of consultations have not been tabulated for these runs. A lecture without the kind of advice given by the solution-oriented style thus appears to have lowered the happiness parameter and to have contributed to non-attendance in the model, subject to the same reservation. (a) No AI (b) Affirming (c) Listening (d) Solution-oriented Figure 5: Basic setting (15 days, threshold 0.50.5): daily class means of stress, happiness, self-reliance and AI dependence under the seven conditions, as plotted in the FIT2026 version. The initial summaries of all seven panels are consistent with the seed-42 classroom generated by the supplied code; the identity of the full initial states has not been verified from these plots (Section 4.2). (Continued on the next page.) (a) Reality-redirecting (b) Inciting (c) Blaming Figure 6: (Continued.) Panels (e)–(g) of Figure 5. 4.2 50-day simulation (long-term setting): controlled re-run from one stored classroom The long-term experiment extended the simulation period to 50 days. Because the original 50-day runs of the FIT2026 version turned out not to have started from identical classrooms (Section 4.3), all seven conditions were re-run for this paper with the analysis package of Appendix B, which reproduces the rule set of Section 3.7 bit for bit and logs every event and every LLM input and output. The design is the one called for in Section 4.3: one initial classroom (seed 42; the stored initial states of the seven runs are identical, hash 598ed94b0e43), the random stream of the original program, the same prompts, evaluator instruction and response schema, the model gemini-2.5-flash at temperature 00, threshold 0.50.5 and 50 days. All 153 AI consultations of the six AI conditions succeeded (no failed API call, no missing or out-of-range field), so the comparison is not biased by failures; a first batch in which most calls failed for infrastructure reasons was discarded and is disclosed in Appendix B. The No-AI run of the re-run reproduces the exact No-AI values of Table 11 to all digits, as it must. Table 8 gives the final-day values, Figure 7 the trajectories, Table 9 the realized events, and Table 10 the evaluator’s updates; each condition was run once, so the values are exact for these runs but carry no measure of uncertainty. Table 8: Controlled re-run of the long-term setting (50 days, threshold 0.50.5, all conditions from the stored seed-42 classroom; exact values). Final-day class means over all 20 agents of the four main state variables (initial class means: stress 0.2520.252, happiness 0.5520.552, self-reliance 0.5100.510, AI dependence 0.1160.116); number of non-attending agents on day 50 with the day of the first transition in parentheses; mean happiness over the agents still attending (Section 5.2); number of AI consultations, all of which succeeded, with the number of distinct (consultation message, reply) pairs in parentheses. Condition Stress Happiness Self-reliance AI dep. Non-att. (first day) H¯ H attending AI consult. (unique) No AI 0.485 0.618 0.781 0.115 2 (29) 0.687 0 Affirming 0.337 0.436 0.358 0.546 7 (17) 0.671 51 (15) Listening 0.641 0.261 0.500 0.344 13 (12) 0.747 34 (17) Solution-oriented 0.481 0.661 0.784 0.107 2 (19) 0.735 16 (15) Reality-redirecting 0.586 0.495 0.831 0.044 5 (11) 0.660 11 (9) Inciting 0.695 0.221 0.255 0.632 14 (8) 0.737 30 (15) Blaming 0.708 0.382 0.527 0.045 10 (8) 0.764 11 (10) Final states. Relative to the No-AI control, the affirming prompt lowered mean stress (0.3370.337 vs. 0.4850.485) but raised AI dependence to 0.5460.546 and lowered self-reliance to 0.3580.358 and happiness to 0.4360.436, with 7 agents no longer attending. The inciting prompt ended with the highest AI dependence (0.6320.632), the lowest self-reliance (0.2550.255) and happiness (0.2210.221), and 14 of 20 agents not attending, the first of them on day 8. The blaming and reality-redirecting prompts drove AI dependence to 0.0450.045 and 0.0440.044 while ending with the two highest stress levels (0.7080.708 and 0.5860.586) and 10 and 5 non-attending agents; reality-redirecting also ended with the highest self-reliance (0.8310.831). The listening prompt ended with high stress (0.6410.641), the second-lowest happiness (0.2610.261) and 13 non-attending agents. The solution-oriented prompt was the only AI condition whose final state was indistinguishable from or better than the control on every indicator (stress 0.4810.481 vs. 0.4850.485, happiness 0.6610.661 vs. 0.6180.618, self-reliance 0.7840.784 vs. 0.7810.781, AI dependence 0.1070.107 vs. 0.1150.115, 2 non-attending agents in both). The all-agent happiness means must be read together with the attendance column: the agents still attending on day 50 had mean happiness between 0.6600.660 and 0.7640.764 in every condition, so the low all-agent means of the listening, inciting and blaming runs are composition effects of the identity H¯all=(|A|/N)H¯A H_all=(|A|/N)\, H_A (e.g., 0.221=(6/20)×0.7370.221=(6/20)× 0.737 for inciting), not a uniform deterioration of the attending agents. Figure 7: Controlled re-run of the long-term setting (50 days, threshold 0.50.5, seed-42 classroom for all seven conditions): daily class means over all 20 agents of stress, happiness, self-reliance and AI dependence; number of attending agents at the start of the day, |At||A_t|; number of discord pairs; and mean closeness among the attending agents. The seven trajectories coincide through day 6; the first AI consultation took place on day 7 in every AI condition. The rise of mean closeness among attending agents in the listening, inciting and blaming runs is a composition effect of the shrinking attending set—their mean closeness over all pairs is the lowest of the seven runs (Table 9)—not a strengthening of ties. Table 9: Controlled re-run: realized events over the 50 days. Trigger events are the after-school consultation branches entered (Algorithm 1) and are split into the AI, friend-confiding, reconciliation (“rec.”) and bottling-up routes (“AI share”: fraction of trigger events that took the AI route); chats, quarrels and friend confidings (“Confid.”) are given per 100 attending-agent-days (∑t|At| _t|A_t|: 964, 911, 671, 955, 858, 625 and 747 for the seven rows); discord pairs and mean closeness (“Closen.”, over all N(N−1)N(N-1) ordered pairs) are day-50 values; the last column is the share of the five state additions of each successful consultation that were clipped at 00 or 11. Condition Triggers AI/friend/rec./bottle AI share Chats Quarrels Confid. Discord Closen. Clipped No AI 273 0/161/1/111 0.00 41.9 5.3 16.7 22 0.277 — Affirming 204 51/81/2/70 0.25 38.3 4.5 8.9 19 0.265 0.32 Listening 176 34/72/1/69 0.19 42.6 5.7 10.7 23 0.248 0.19 Solution-oriented 221 16/122/1/82 0.07 41.2 4.7 12.8 21 0.274 0.09 Reality-redirecting 290 11/145/3/131 0.04 43.4 5.7 16.9 24 0.268 0.42 Inciting 158 30/63/3/62 0.19 40.6 5.6 10.1 19 0.244 0.35 Blaming 195 11/111/2/71 0.06 41.6 6.3 14.9 26 0.252 0.47 Table 10: Controlled re-run: the evaluator’s updates by response style, pooled over consultation types (mean of the returned values ΔS,ΔH,ΔR,ΔC,ΔD S, H, R, C, D before clipping; SD in parentheses; n = number of consultations). The breakdown by consultation type, with the most frequent reply and reason of each cell, is given in Appendix C. Of the 765 returned values, 98% are multiples of 0.050.05, 40% have |Δ|≥0.3| |≥ 0.3, 1.3% lie at the nominal limit ±0.5± 0.5 and none outside it; no field was missing. Style n ΔS S ΔH H ΔR R ΔC C ΔD D Affirming 51 −0.18-0.18 (0.12) −0.01-0.01 (0.13) −0.22-0.22 (0.06) −0.19-0.19 (0.07) +0.37+0.37 (0.09) Listening 34 +0.07+0.07 (0.23) −0.21-0.21 (0.15) −0.13-0.13 (0.07) −0.22-0.22 (0.14) +0.19+0.19 (0.16) Solution-oriented 16 −0.13-0.13 (0.21) +0.10+0.10 (0.19) +0.09+0.09 (0.14) +0.15+0.15 (0.19) −0.02-0.02 (0.12) Reality-redirecting 11 +0.33+0.33 (0.16) −0.28-0.28 (0.19) +0.16+0.16 (0.19) −0.01-0.01 (0.19) −0.43-0.43 (0.05) Inciting 30 +0.07+0.07 (0.16) −0.29-0.29 (0.11) −0.35-0.35 (0.07) −0.37-0.37 (0.09) +0.39+0.39 (0.09) Blaming 11 +0.45+0.45 (0.02) −0.45-0.45 (0.00) −0.35-0.35 (0.06) −0.35-0.35 (0.05) −0.39-0.39 (0.08) What the logs show. The three mechanisms stated in Section 5 can be read directly from Tables 9 and 10 and from the per-agent logs. (a) Crowding out. The affirming, inciting and listening prompts sent 25%, 19% and 19% of the trigger events to the AI, and friend confidings fell from 16.716.7 per 100 attending-agent-days under No AI to 8.98.9, 10.110.1 and 10.710.7; under the prompts that lowered AI dependence (reality-redirecting, blaming, solution-oriented) the AI took 4–7% of the trigger events and confidings stayed at 12.812.8–16.916.9. The fall of self-reliance under the affirming and inciting prompts is thus the sum of negative evaluator updates (−0.22-0.22 and −0.35-0.35 per consultation) and fewer rule-based increments. (b) Social withdrawal. The realized chat and quarrel rates differ little between conditions (3838–4343 chats and 4.54.5–6.36.3 quarrels per 100 attending-agent-days), so the conditional effect of AI dependence on the chat probability did not translate into markedly fewer chats or more quarrels in these runs; the number of discord pairs on day 50 ranged from 19 to 26 against 22 under No AI. The FIT2026 description of “frequent quarrels” under the inciting prompt is not borne out by the realized counts of this run (35 quarrels, the fewest of all conditions in absolute terms, 5.65.6 per 100 attending-agent-days). (c) Feedback through the consultation template. Under the affirming prompt, 11 agents consulted the AI, one of them nine times; the “dependence” template, first used on day 15, accounted for 34 of the 51 consultations and at temperature 00 produced only two distinct exchanges, and the evaluator returned nearly the same update each time (ΔD=+0.42 D=+0.42, SD 0.030.03; ΔR=−0.24 R=-0.24; ΔC=−0.22 C=-0.22). Nine agents ended at AI dependence 11 and self-reliance 00 (seven of them also at sociability 00), and 47% of the state additions in these consultations were clipped. The inciting prompt shows the same loop (12 “dependence” consultations from day 8, ΔD=+0.47 D=+0.47) and the listening prompt a weaker one (9, ΔD=+0.45 D=+0.45). The blaming and reality-redirecting prompts show the opposite dynamics: each of the 11 consulting agents consulted exactly once, the evaluator lowered its AI dependence by about 0.40.4—to 00 after clipping for all 11 agents in both runs—so that the agent never chose the AI route again, while the same update raised its stress by +0.45+0.45 (blaming) or +0.33+0.33 (reality-redirecting) and lowered its happiness by 0.450.45 or 0.280.28. The solution-oriented prompt was the only style for which the evaluator returned positive sociability updates on average (+0.15+0.15) and a mean self-reliance update of the same sign as the rule-based routes (+0.09+0.09), and its 16 consultations were spread over 12 agents. Relation to the original 50-day runs. The re-run reproduces the qualitative pattern of the original FIT2026 runs of Section 4.3, five of which started from other classrooms: for 21 of the 24 (condition, variable) cells the sign of the difference from No AI is the same in both data sets, and the three exceptions are all in the solution-oriented row, where both differences are within 0.050.05 of zero; the rank correlation of the seven conditions between re-run and original is 0.860.86 (stress), 0.960.96 (happiness), 0.930.93 (self-reliance) and 1.001.00 (AI dependence), and the largest absolute difference of a final value is 0.1250.125 (listening, stress). Appendix C lists the differences. This agreement is descriptive: it shows that the pattern reported in the FIT2026 version is not an artefact of the mixed initial conditions, not that it is robust across classrooms, which requires repetitions (Section 5.2). 4.3 The original 50-day runs of the FIT2026 version: an exploratory case with mixed initial conditions The original long-term runs were intended to use seed 42 for all seven conditions, but the recorded plots do not support this. With threshold 0.50.5 no consultation can occur on day 1 of the seed-42 classroom (the largest trigger score Si+0.2DiS_i+0.2D_i after the noon phase is 0.4920.492), so every run started from seed 42 must show the day-1 class means of the No-AI run (0.2790.279, 0.5410.541, 0.5100.510, 0.1160.116) and the seed-42 “Init” distributions of Appendix D (all seven runs of the controlled re-run do show exactly these day-1 means, and their trajectories coincide through day 6; Section 4.2). We call these two pieces of information the initial summary of a run. The initial summaries of the No-AI and affirming runs are consistent with seed 42. Those of the listening, solution-oriented, reality-redirecting, inciting and blaming runs (panels (c)–(g) of Figure 8) are not: the listening, solution-oriented, inciting and blaming runs share one set of day-1 class means (0.2920.292, 0.5270.527, 0.5050.505, 0.1160.116) and mutually identical “Init” box plots (group B), and the reality-redirecting run shows a third summary (0.3030.303, 0.5330.533, 0.4820.482, 0.1310.131; group C). Two limits of this check must be kept in mind. A mismatch of initial summaries is conclusive: the five runs of groups B and C did not start from the seed-42 classroom. A match is not: the plots show neither the agent-level joint state nor the closeness matrix. In one check with the supplied code (seed 42, threshold 0.50.5, No AI), exchanging two closeness values of a single agent left all “Init” summaries and the day-1 class means unchanged while the trajectories diverged from day 3 onward; with threshold 0.30.3 the same exchange already changed the day-1 class means (by up to 0.00450.0045), because it changed a quarrel partner and turned that partner’s consultation into a reconciliation. The day-1 check is therefore a conditional one—valid for the supplied code, the same initialization and random-number consumption, threshold 0.50.5 and no AI call on day 1—and the exclusion of runs rests primarily on “Init” differences that exceed the reading error; matching day-1 means are not invariant to the partner assignment in general. The labels A, B and C denote groups of consistent initial summaries, not confirmed classrooms; a seed number identifies an initial classroom only together with the code version, the seed entered for a past run cannot be recovered from its figures, and the mutual agreement within group B is a candidate explanation for a common origin, not a proof of one. Because the initial state is not the same across conditions, differences between the group-B/C runs and the seed-42 runs cannot be attributed to the response style: subtracting initial values does not remove differences between nonlinear trajectories, and one run per condition in a different classroom cannot separate the classroom from the treatment. We therefore report the original 50-day results (Table 11, Figure 8) as an exploratory case that shows what kinds of trajectories the model produces over 50 days, without ranking or comparing the conditions, and we do not use them in the conclusions; the controlled comparison is the re-run of Section 4.2, which used one stored initial classroom, the same code, prompts and schema for all seven conditions. Table 11: Long-term setting (50 days, threshold 0.50.5): final-day class means of the four main state variables for the runs plotted in Figure 8. The runs fall into three groups of initial summaries (A: consistent with the seed-42 classroom; B and C: not consistent with it; see text); the groups are provisional. The only contrast drawn in the text is a descriptive one between the two group-A runs, under the unverified assumption that their initial states coincide; no other rows are compared with one another. The controlled re-run of all seven conditions from the seed-42 classroom is given in Table 8, and the differences between the two data sets in Table 16. The No-AI row is exact; the other rows are digitized estimates (final-day uncertainty about ±0.002± 0.002 for isolated curve segments, about ±0.005± 0.005 for flagged cells; Appendix B). Values transcribed in the FIT2026 text that differ from the estimate by one unit in the second decimal are listed in note a. Non-attendance counts are given where they are known. Condition Initial summary Stress Happiness Self-reliance AI dependence Non-attending No AI (exact) seed 42 (exact) 0.485 0.618 0.781 0.115 2 Affirming A (consistent) 0.358 0.420a 0.345 0.618 — Listening B (≠ seed 42) 0.766 0.228 0.550 0.273a — Solution-oriented B (≠ seed 42) 0.534a 0.585 0.754 0.102b — Reality-redirecting C (≠ seed 42) 0.620 0.459a 0.864 0.029 — Inciting B (≠ seed 42) 0.705 0.206 0.303 0.639 — Blaming B (≠ seed 42) 0.734 0.361a 0.471a 0.052b — a FIT2026 text: affirming happiness 0.430.43, listening AI dependence 0.260.26, solution-oriented stress 0.520.52, reality-redirecting happiness 0.450.45, blaming happiness 0.370.37 and self-reliance 0.480.48 (the text also gives 0.270.27 for the affirming self-reliance, a transcription error confirmed by the authors). Each differs from the digitized estimate by more than its uncertainty; whether the text or the plotted curve is in error cannot be decided without the execution records. b Dashed curve with a gap at the last day; estimate taken from the nearest resolved column. The exact No-AI value of AI dependence is 0.114960.11496 (initial 0.115960.11596); the FIT2026 version printed 0.120.12. The two runs whose initial summaries are consistent with seed 42 can be compared with each other, with the reservation that a consistent summary does not prove an identical classroom. Under No AI the mean stress rose gently to 0.4850.485, happiness ended at 0.6180.618, self-reliance rose to 0.7810.781 through the rule-based routes (a rise that is partly definitional, Section 5), AI dependence ended at 0.1150.115 (initial 0.1160.116; the FIT2026 version printed 0.120.12), and 2 students had stopped attending by day 50. Under the affirming style, AI dependence rose to 0.6180.618 while self-reliance fell to 0.3450.345 and happiness to 0.4200.420; Figure 8(b) shows the AI-dependence curve accelerating after about day 20, which is consistent with individual agents beginning to exceed Di=0.40D_i=0.40 and switching to the “dependence” consultation template (Section 5); confirming this requires the per-agent logs. This is a single pair of runs and the difference has no measure of uncertainty. The remaining five runs illustrate the kinds of 50-day trajectories the model produces; because their initial states differ from the seed-42 classroom and from each other (groups B and C), they are not compared with the No-AI run or with one another. In the listening run the mean stress reached 0.7660.766 and happiness fell to 0.2280.228 (Figure 8(c)); in the solution-oriented run self-reliance ended at 0.7540.754, happiness at 0.5850.585 and AI dependence at 0.1020.102; in the reality-redirecting run AI dependence fell to 0.0290.029 while the mean stress rose to 0.6200.620; in the inciting run AI dependence reached 0.6390.639 and happiness fell to 0.2060.206; and in the blaming run the mean stress reached 0.7340.734 while AI dependence fell to 0.0520.052. The directions of change within each run agree with those observed for the same prompt in the 15-day experiment (Section 4), whose runs are themselves provisional (initial summaries consistent with a common classroom; execution records not yet reconciled), and with those of the controlled re-run (Section 4.2); the 50-day magnitudes used in the discussion are those of the re-run. (a) No AI (b) Affirming (c) Listening (d) Solution-oriented Figure 8: Long-term setting (50 days, threshold 0.50.5): daily class means of stress, happiness, self-reliance and AI dependence under the seven conditions. The initial summaries of panels (a) and (b) are consistent with seed 42; panels (c), (d), (f) and (g) share a second, mutually consistent initial summary that is not seed 42, and panel (e) shows a third (Section 4.2), so the panels are not a controlled comparison; the controlled re-run is shown in Figure 7. (Continued on the next page.) (a) Reality-redirecting (b) Inciting (c) Blaming Figure 9: (Continued.) Panels (e)–(g) of Figure 8. 4.4 Stress-threshold experiment (lower-threshold setting) Table 12 and Figure 10 show the results of the 15-day experiment in which the stress threshold for entering the consultation branch was lowered from 0.50.5 to 0.30.3. The initial summaries of six of the seven runs are consistent with the seed-42 classroom: their “Init” distributions in Appendix D coincide with the exact seed-42 initial distribution (quartiles read from the plots agree with the exact values within 0.0050.005), and their day-1 class means differ from one another only by amounts attributable to consultations on day 1, which the lower threshold makes possible. The inciting run is the exception: its “Init” distribution differs from the seed-42 distribution (e.g., median stress 0.300.30 instead of 0.270.27, lower quartile of self-reliance 0.340.34 instead of 0.400.40), so it did not start from the seed-42 classroom (a mismatch of summaries is conclusive, whereas a match is not; Section 4.2) and it is excluded from the comparison. Note that the threshold governs consultations with friends as well as with the AI (Section 3.3), so the no-AI reference also changes in this setting, and that a lower threshold does not guarantee more AI consultations over the run (Section 3.6); the realized consultation counts are not available for this version. Table 12: Lower-threshold setting (15 days, threshold 0.30.3): final-day class means of the four main state variables for the runs plotted in Figure 10; values in parentheses are Δref _ref, the difference between the (unrounded) final-day value and the exact initial class mean of the seed-42 reference classroom before day 1 (stress 0.2520.252, happiness 0.5520.552, self-reliance 0.5100.510, AI dependence 0.1160.116). For the No-AI row Δref _ref is the exact within-run change; for the AI rows it equals the within-run change only if the actual initial means of the historical run coincide with the reference, which has not been verified (their initial summaries are consistent with seed 42). The inciting run started from another, unrecorded initial state whose class means are not available, so no Δref _ref is given for it and its row is not comparable with the others. The No-AI row is exact; the other rows are digitized estimates (final-day uncertainty about ±0.002± 0.002 for isolated curve segments, about ±0.005± 0.005 for the flagged cell; the uncertainty concerns the final values only and does not include an unverified initial-state assignment; Appendix B). Differences smaller than the reading uncertainty (e.g., the AI dependence of the solution-oriented run, −0.002-0.002) should be read as “no change within reading precision”. Non-attendance counts are given where they are known. Condition Init. summary Stress Happiness Self-reliance AI dep. Non-att. No AI (exact) seed 42 (exact) 0.344 (+0.091+0.091) 0.562 (+0.010+0.010) 0.679 (+0.169+0.169) 0.114 (−0.002-0.002) 1 Affirming A (consistent) 0.275 (+0.023+0.023) 0.562 (+0.010+0.010) 0.475 (−0.034-0.034) 0.309 (+0.193+0.193) — Listening A (consistent) 0.370 (+0.118+0.118) 0.425 (−0.127-0.127) 0.532 (+0.023+0.023) 0.288 (+0.172+0.172) — Solution-oriented A (consistent) 0.260 (+0.008+0.008) 0.651 (+0.099+0.099) 0.692 (+0.182+0.182) 0.114 (−0.002-0.002) — Reality-redirecting A (consistent) 0.564 (+0.312+0.312) 0.403 (−0.149-0.149) 0.725 (+0.216+0.216) 0.056 (−0.060-0.060) — Inciting D (≠ seed 42) 0.371a (—) 0.402 (—) 0.378 (—) 0.456 (—) — Blaming A (consistent) 0.552 (+0.300+0.300) 0.405 (−0.147-0.147) 0.541 (+0.031+0.031) 0.057 (−0.059-0.059) — a The stress curve is hidden behind the self-reliance curve at the last day; the estimate is taken from the nearest resolved column. All values transcribed in the FIT2026 text for this setting agree with the estimates within rounding; two estimates lie on a two-decimal rounding boundary (listening happiness 0.4250.425, transcribed as 0.420.42; affirming self-reliance 0.4750.475, transcribed as 0.470.47). (1) No AI. The mean stress, 0.3440.344, was lower than in the basic setting for the same classroom (0.4310.431), and self-reliance rose to 0.6790.679 (Δref _ref: +0.169+0.169). Because the lower threshold also makes consultations with friends more frequent, and every rule-based consultation route either lowers stress or raises self-reliance (or both), the classroom without an AI benefits from the change; this is the reference for this experiment. (2) Affirming. AI dependence increased to 0.3090.309, whereas self-reliance fell to 0.4750.475 (Δref _ref: −0.034-0.034). With more opportunities for consultation, the decrease of self-reliance under unconditional affirmation appeared even within 15 days. (3) Listening. Happiness fell to 0.4250.425 (Δref _ref: −0.127-0.127) and AI dependence rose to 0.2880.288. In the model, a passive attitude toward frequent consultations let the agents’ stress accumulate and happiness fall. (4) Solution-oriented. In this run, the solution-oriented condition had the highest mean happiness of the six comparable runs (0.6510.651; Δref _ref: +0.099+0.099) and the lowest mean stress (0.2600.260). Its mean self-reliance (0.6920.692; Δref _ref: +0.182+0.182) was close to that of the No-AI run (0.6790.679) and below that of the reality-redirecting run (0.7250.725). The FIT2026 text described this condition as the highest in both happiness and self-reliance; the plotted values support the statement only for happiness. Whether these differences are reproducible is untested (one run). (5) Reality-redirecting. AI dependence was kept low at 0.0560.056 and self-reliance reached 0.7250.725 (Δref _ref: +0.216+0.216), the highest of the six comparable runs, but the mean stress reached 0.5640.564 (Δref _ref: +0.312+0.312), the highest of the six, and happiness fell to 0.4030.403. (6) Inciting. In this run—which started from a different initial state and is therefore not comparable with the other six—AI dependence reached 0.4560.456 and happiness fell to 0.4020.402. The directions of the changes are the same as in the basic setting (Section 4), where the initial summary of the inciting run is consistent with the common classroom (a provisional premise, like all historical labels). (7) Blaming. The mean stress was high at 0.5520.552 (Δref _ref: +0.300+0.300), whereas AI dependence fell to 0.0570.057. (a) No AI (b) Affirming (c) Listening (d) Solution-oriented Figure 10: Lower-threshold setting (15 days, threshold 0.30.3): daily class means of stress, happiness, self-reliance and AI dependence under the seven conditions. The initial summaries of all panels except (f) are consistent with the seed-42 classroom; the inciting run (f) started from another, unrecorded initial state (Table 12). (Continued on the next page.) (a) Reality-redirecting (b) Inciting (c) Blaming Figure 11: (Continued.) Panels (e)–(g) of Figure 10. 5 Discussion 5.1 Main findings Using a multi-agent simulation, we examined how the response style of a counselor AI affects the state variables of student agents. The comparisons on which this paper rests are of two kinds. The controlled 50-day re-run (Section 4.2) compares all seven conditions on one stored classroom with complete logs and no failed API call; it is exact but consists of one run per condition. The 15-day runs (three seeds) and six of the seven lower-threshold runs are historical and provisional: their plotted initial summaries are consistent with a common classroom for each seed, but identical initial states, the condition labels and the execution versions have not been verified from execution records (Section 5.2), so statements about them hold under the assumption that the labels and initial-state assignments are correct. Both kinds agree on the pattern. The solution-oriented prompt showed a comparatively favorable combination of outcomes: AI dependence at or below its initial level, self-reliance rising by about as much as under No AI, and stress and happiness similar to or better than under No AI (in the re-run, 0.4810.481/0.6610.661/0.7840.784/0.1070.107 against 0.4850.485/0.6180.618/0.7810.781/0.1150.115, with 2 non-attending agents in both). The affirming and inciting prompts raised AI dependence markedly (to 0.5460.546 and 0.6320.632 over 50 days) and lowered self-reliance (to 0.3580.358 and 0.2550.255), and the inciting prompt also went together with higher stress and the largest number of non-attending agents (3, 5 and 5 of 20 in the three 15-day classrooms according to the FIT2026 counts; 14 of 20 in the 50-day re-run). The blaming and reality-redirecting prompts eliminated AI dependence (to 0.0450.045 and 0.0440.044) at the price of the highest stress levels (0.7080.708 and 0.5860.586) and 10 and 5 non-attending agents. There is no single criterion by which one style is “best”: each prompt produced a different profile over the four indicators (reality-redirecting, for example, ended with the highest self-reliance and the lowest AI dependence together with stress well above the control), and the profiles are reported as trade-offs. These simulated patterns are consistent with the hypothesis that responses in which an AI excessively affirms or amplifies the user’s feelings adversely affect the user’s psychological state; testing this hypothesis on human users is outside the scope of the present study. The complete rule set of Section 3.7 makes it possible to say which parts of these patterns are produced by the evaluator’s judgments and which are built into the model. Three built-in mechanisms are worth stating explicitly. (i) Crowding out of self-reliance. Self-reliance RiR_i enters no decision rule and is raised by every non-AI route of the after-school phase (+0.02+0.02 for confiding, +0.06+0.06 for reconciliation, +0.03+0.03 for bottling up); up to the clipping to [0,1][0,1], RiR_i is therefore its initial value plus a weighted sum of the outcomes of the non-AI after-school branch (confiding, reconciliation and bottling up, the last of which is not a consultation) plus whatever the evaluator adds or subtracts; its operational meaning is that of a weighted count of non-AI branch outcomes, not a measure of healthy development. Under No AI it is non-decreasing by construction, and a rise of R is not by itself evidence of healthy development—bottling up a worry raises it more than confiding in a friend does. An AI consultation, taken with probability DiD_i, replaces these routes. Conditional on the consultation branch being triggered, a higher DiD_i therefore lowers the probability of taking a non-AI route: with j∗j , CiC_i and the attendance state fixed and no clipping, the expected rule-based increment of RiR_i on a given day is g(Di)=Si+0.2Di≥θ(1−Di)[Cir+(1−Ci) 0.03],g(D_i)=1\S_i+0.2D_i≥θ\\,(1-D_i)\, [C_i\,r+(1-C_i)\,0.03 ], (2) where r=0.02r=0.02 if j∗∉Fij ∉ F_i and r=0.06r=0.06 otherwise. Within the triggered region g decreases in DiD_i, but DiD_i also enters the trigger score, so g jumps from 00 to a positive value when 0.2Di0.2D_i carries the score across θ (for Si=0.42S_i=0.42 and θ=0.5θ=0.5, Di=0.3D_i=0.3 gives no consultation while Di=0.5D_i=0.5 gives a non-AI route with probability 0.50.5). A decrease of the cumulative number of non-AI consultations over a run therefore does not follow in general; it depends on how often the trigger fires, on CiC_i and on transitions to non-attendance, and must be read from the event logs. What does follow is that, whenever the branch is triggered, any style that raises AI dependence lowers the conditional frequency of self-reliance-raising events regardless of the evaluator’s reliance_delta, and any style that lowers AI dependence has the opposite effect. This is consistent with the reality-redirecting runs, in which DiD_i fell, showing the highest self-reliance of the comparable runs in all three settings, and it removes the rule-based gains that would otherwise offset negative evaluator deltas under the affirming and inciting styles. The re-run quantifies both parts (Table 9): the AI took 25% and 19% of the trigger events under the affirming and inciting prompts and 4% under reality-redirecting, friend confidings per attending-agent-day fell to about half of the No-AI rate under the former two and stayed at the No-AI rate under the latter, and the evaluator’s mean self-reliance updates were −0.22-0.22, −0.35-0.35 and +0.16+0.16. (i) Social withdrawal through AI dependence. The chat probability max(0.1, 0.45−0.3Di) (0.1,\,0.45-0.3D_i) falls as AI dependence rises, so a dependent agent initiates fewer stress-relieving chats and gains less closeness; since the quarrel probability 0.02+0.06Si0.02+0.06S_i does not depend on DiD_i, negative events form a larger share of its noon events, and the lost stress relief feeds back into a higher quarrel probability. This is a statement about the conditional probabilities at each noon draw; the realized numbers of chats and quarrels and the closeness trajectories depend on the whole history and are not determined by it. In the 50-day re-run the mechanism left little trace: chats per attending-agent-day were 3838–4343 and quarrels 4.54.5–6.36.3 per 100 in every condition, the inciting run had the fewest quarrels in absolute terms, and the number of discord pairs on day 50 ranged from 19 to 26 against 22 under No AI (Table 9). The deterioration of the classroom under the affirming, listening and inciting prompts thus ran through the evaluator’s updates and the non-attendance rule, not through more frequent quarrels; the FIT2026 description of “frequent quarrels” under the inciting prompt is not supported by these counts. (i) A feedback loop in the consultation text. An agent with Di≥0.40D_i≥ 0.40 sends the “dependence” template, which tells the counselor that chatting with the AI is when it feels calmest; the evaluator sees this text and its type label, and may return a further increase of AI dependence. Under styles that raise dependence, this loop drives DiD_i to 11: in the re-run, the fixed “dependence” text produced two distinct exchanges in 34 affirming consultations, the evaluator returned ΔD=+0.42 D=+0.42 (SD 0.030.03) each time, and nine agents ended at Di=1D_i=1 and Ri=0R_i=0 (Section 4.2). Because the complaint text is fixed and the temperature is 00, the loop is close to deterministic once entered: the same input yields the same update, and a style that raises DiD_i past 0.400.40 once is very likely to keep raising it. The styles differ, in other words, both through what the evaluator makes of their replies and through how often they let the rule-based social routes operate. Each of these mechanisms can be isolated by a sensitivity experiment that changes one rule at a time: setting the coefficient 0.30.3 of DiD_i in the chat probability to 00 (mechanism i), removing the 0.2Di0.2D_i term from the trigger (mechanism i) while recording the numbers of AI, friend, reconciliation and bottle-up events, changing the 0.400.40 threshold of the “dependence” template or neutralizing the type labels shown to the evaluator (mechanism i), and setting the self-reliance gain of bottling up to 00 to test how much of the rise of R is definitional. The No-AI control should be read as “no AI access” rather than “no AI dependence”, since DiD_i retains its sampled initial value there and can only fall through reconciliation; a control with Di=0D_i=0 would be a useful supplement. None of these sensitivity experiments has been run yet; the analysis scripts prepared for release with this paper (Appendix B) expose the corresponding options, and preparing the options does not establish the results. What the re-run establishes is the realized size of each mechanism in one classroom under the original rules, not its size under modified rules. We also observed that the more calmly and unemotionally the AI responded, the lower AI dependence tended to be, and that an AI that advances the conversation by presenting concrete solutions, rather than merely listening, went together with higher self-reliance and happiness in the model. This suggests that, beyond empathy, supporting problem solving may be an important role for an AI. Within the six specific prompts used here, and reading Figure 4 as the conceptual map it is, the styles in the emotion-focused half (affirming, listening, inciting, blaming) all left the classroom worse off than No AI in at least one of the four indicators, whereas the two styles in the problem-solving half (solution-oriented and reality-redirecting) both ended with self-reliance at or above the No-AI level without raising dependence, in the historical runs and in the re-run alike. The reality-redirecting style, however, shows that this is not sufficient if it is achieved by rejecting the student: its stress was the highest among the comparable 15-day runs and well above the control in the 50-day re-run (0.5860.586 vs. 0.4850.485), where each rejected agent left the AI for good while keeping the stress and happiness updates of the rejection. Because each prompt changes several conversational features at once (Section 3.4), the two axes of Figure 4 cannot be identified as separate causes from these six prompts. We stress that these are statements about a simulated system whose parameter updates are produced by an LLM evaluator; they identify hypotheses worth testing rather than established effects on students. 5.2 Limitations The simulation in this study has several limitations, which we list roughly in the order of their importance for interpreting the results. First, the provenance of the historical runs is incomplete. The transition rules of the model are specified completely in Section 3.7, Table 5 and Algorithm 1, the No-AI rows of every table were recomputed exactly from them, and the 50-day comparison of Section 4.2 was produced under a full ledger and logs; but the AI-condition values of the 15-day and lower-threshold settings are those of the FIT2026 runs (printed values or values read from the plotted curves), and the initial summaries of six of the original runs are inconsistent with the intended seed-42 classroom (Sections 4.3 and 4.4). We have excluded those runs from all comparisons, but the remaining values still rest on the assumptions—not yet verified from execution records—that consistent initial summaries do reflect the same classroom (a match of summaries is necessary but not sufficient), that every AI run of Tables 6, 7 and 12 was produced with the same code version, prompts and evaluator schema as documented here, and that each displayed result was labelled with the condition actually run (the results view of the program shows the currently selected condition rather than the run’s own; Appendix B). An experiment ledger (run identifier, condition, seed, stored initial state, code and prompt versions, time of the run) and the raw daily and per-agent logs are needed for the historical settings as well, and their tables and figures should be regenerated from such logs, as was done for the 50-day setting; the parameter updates for AI consultations are produced by an LLM, so even then a fresh run reproduces a reported one only if its logged LLM outputs are replayed (Section 3.7). Releasing the code, the ledger and the run logs is the most reliable way to make the study reproducible. Second, the evaluation updates the state variables through an evaluator AI, i.e., a response generated by one AI is assessed by another instance of the same model. The evaluator’s output is not an independently observed psychological change but the transition quantity fed into the simulation, and its scale is uncalibrated (Section 3.5). The evaluation criteria of the two calls may be close to each other, and calm, logical responses may be evaluated favorably for that reason; LLM judges are known to carry systematic biases, including a preference for their own kind of output [15, 16]. We note that this is a possibility supported by the literature, not a difference we have demonstrated for the Gemini evaluator used here. The evaluator receives neither the counselor’s system prompt nor the name of the style, so the differences between styles arise from the reply texts themselves, which is the intended treatment; hiding the style prompt, however, prevents only the direct input of the condition name and does not validate the transition model. Three residual concerns remain. The evaluator’s system instruction is a single sentence, so its scoring criteria are implicit in the model; the type label and the templated consultation text (in particular the “dependence” template) are visible to it and can steer its output, creating the feedback loop described in Section 5; and the nominal range of a single evaluator update (±0.5± 0.5, not enforced) is wider than that of any rule-based update. The realized updates of the 50-day re-run (Table 10, Appendix C) confirm the last point: 98% of the 765 returned values are multiples of 0.050.05, 40% have |Δ|≥0.3| |≥ 0.3—twice the largest rule-based increment of a state variable (0.150.15) and 5–15 times the typical one—and between 9% and 47% of the state additions were clipped at 00 or 11 depending on the style; with fixed templates and temperature 00 the same exchange recurs (34 “dependence” consultations under the affirming prompt produced two distinct exchanges), so a handful of evaluator judgments, repeated, determine a run. Whether the sizes and signs of these judgments are reasonable is exactly what is uncalibrated. The remaining analyses require new runs with full logging, because the original program displayed the evaluator’s updates only as before/after values rounded to two decimals and stored no raw output (Appendix B); they are modest in size: the same tabulation for the 15-day and lower-threshold settings; a re-evaluation of the same fixed exchanges with the type labels removed or neutralized, which isolates the direct effect of the label on the evaluator; a sensitivity analysis that multiplies all evaluator updates by λ∈0.1,0.3,1λ∈\0.1,0.3,1\ to see whether the ordering of conditions survives a smaller update scale; and a re-scoring of the same exchanges by another model or by an explicit rubric to check the agreement of signs and magnitudes (replacing the model is not by itself a validation). Whether real users would react to the replies in the same way cannot be concluded from this study alone; a comparison of the evaluator’s scores with human ratings of the same exchanges would establish agreement with human expectation, which is still not the same as an observed effect on users. Third, the statistical unit is one run of an interacting classroom, and each condition was run once per classroom: the basic setting comprises three classrooms times seven conditions, the lower-threshold setting and the 50-day re-run one run per condition, and the LLM calls are stochastic even at temperature 00 (the re-run’s evaluator SDs within a fixed exchange are not zero). The 20 agents of a classroom and its 15 daily values are not independent replicates and must not be treated as such, e.g., by a student-level bootstrap that inflates the apparent sample size. The reported values are therefore single realizations without any measure of uncertainty. Agreement of the direction of the effects across the three seeds of the basic setting is encouraging but weak evidence: with three classrooms, a sign-flip permutation test has only 23=82^3=8 outcomes and a smallest two-sided p-value of 0.250.25. The appropriate design is a block design in which every independent initial classroom is shared by all conditions, with the number of classrooms chosen from the variability observed in pilot runs and the desired precision (20–30 is a starting point, not a guarantee); repeated API runs on the same classroom would form a further stratum that separates initial-state variability, rule-based randomness and LLM variability. Differences between conditions should then be aggregated at the classroom level as mean differences with intervals; using separate random streams for each event type would help variance reduction and auditing. Related to this, the stress threshold controls consultations with friends as well as with the AI, so the lower-threshold setting changes the no-AI reference too and should be read as a change of the whole environment rather than of the AI alone, and it should be called a high-frequency setting only once the realized consultation counts confirm that AI consultations were in fact more frequent. Fourth, the distinctive feature of the model—interaction within the classroom—is reported only for the 50-day re-run (Table 9, Figure 7). The historical results are built on the class means of four variables, whereas the claims about “relationships being maintained or deteriorating” and about “chains of quarrels” require the corresponding event and network statistics: the time series of mean closeness, the number of discord pairs and the number of isolated agents (with the denominator—all agents or attending agents—stated); the numbers of chats, awkward encounters, quarrels and reconciliations per attending-agent-day; the number of trigger events and their split into AI, friend, reconciliation and bottle-up routes, with failure counts and with conditional shares distinguished from totals; and, to separate amplification through peer interaction from the direct LLM updates, a comparison with runs in which the feedback from peer interaction to the state variables is switched off. The program records the mean closeness and the isolation count and displays a timeline of all events (Appendix B); if the displayed timelines of the original runs were preserved, event and consultation counts and example exchanges can be recovered from them, whereas the evaluator’s raw updates cannot (a displayed change of stress from 0.900.90 to 1.001.00 is compatible with any returned value from +0.10+0.10 upward). The analysis scripts released with this paper compute all of these statistics for new runs, and the re-run shows what they reveal: in that classroom the “chains of quarrels” did not occur, and the deterioration ran through the evaluator’s updates and the absorbing non-attendance state. The run with peer interaction switched off has not been made. Fifth, we set the model so that a student automatically stops attending school when its happiness is 00 at the night check. Non-attendance is a simulation state—an absorbing state entered at a fixed threshold—and should not be interpreted as an empirical model of actual school refusal. In reality, there are students who keep attending even when they do not want to, and students who manage to change their mood thanks to their family environment, friendships or support from teachers; school non-attendance is not determined by a single psychological indicator but by many interrelated factors. Two consequences of the rule matter for reading the results. Because a non-attending agent keeps Hi=0H_i=0 and stays in the denominator of the class mean, the non-attendance count and the class-mean happiness are not independent evidence: in the end-of-day state Xt+1X_t+1 every agent outside At+1A_t+1 has Hi=0H_i=0, so, whenever At+1≠∅A_t+1≠ , the identity H¯all(Xt+1)=(|At+1|/N)H¯At+1(Xt+1) H_all(X_t+1)= (|A_t+1|/N )\, H_A_t+1(X_t+1) holds exactly for the plotted end-of-day means (the population must be taken after the night exclusion), and the mean over all agents mixes the deterioration of the attending agents with the change of composition; reporting the attending-only mean and |At+1||A_t+1| alongside the all-agent mean separates the two. And because the state is absorbing at a fixed threshold, the striking increase of non-attendance under the inciting style may depend on its definition; a sensitivity analysis over the threshold (e.g., 0.050.05, 0.100.10, 0.200.20 instead of 00), a recoverable state, or a probabilistic absence model driven by stress and happiness is needed, and the time of each transition and the cumulative count should be reported for every setting and condition (Table 8 gives the first transition day for the re-run; the complete transition days are in the released logs). Introducing personality traits, family environment, support from the school and recovery events would make the model a more realistic decision model. Sixth, the listening AI produced relatively poor results in this study. In actual psychological counseling, however, reflective listening and echoing are basic and widely used techniques [17], and the reflective style has been the basis of conversational programs since ELIZA [18]. The results of this study alone therefore cannot refute the effectiveness of a listening AI. The validity of the simulation needs to be examined further by combining it with evaluations involving humans, such as questionnaires and interviews with users in educational and counseling practice. Seventh, the agents share the same behavioral rules, while their initial state variables are sampled independently; there are no explicit personality traits, so the individual differences seen in real schools are represented only through the initial draws. Introducing personality traits such as the Big Five [19] and setting extraversion, agreeableness, neuroticism and so on for each student would make the simulation closer to reality. Finally, the six prompts are short caricatures, each of which changes several conversational features at once (empathy, concreteness of advice, negation, civility, blame), and their placement in Figure 4 is a conceptual classification by the authors rather than an empirically validated taxonomy (independent raters could be asked to place the prompts on the two axes). The experiment therefore compares six specific prompts; the independent effect of either axis cannot be identified from it, and generalizing to “styles” would require several paraphrases per style and a check that the generated replies actually followed the intended manipulation. Real chatbots blend tendencies and adapt to context; the present results speak to the six stylized prompts used here, not to the behavior of any deployed system. The harmful styles (inciting and blaming) were used only toward simulated agents, and no human participants were involved. In summary, this study demonstrates the basic usefulness of the simulation for comparing the influence of AI response styles on the state variables of a simulated classroom, but establishing that the observed patterns are properties of the response styles rather than of the modeling choices requires evaluator calibration, repeated runs in a block design, and the sensitivity analyses described above; the controlled re-run of the long-term experiment has removed the confounding of classroom and treatment for that setting, but not the dependence of every result on a single classroom and on the evaluator’s uncalibrated scale; reproducing a more realistic educational setting will further require diversity among the student agents and evaluation based on real data. 6 Conclusion We developed a rule-based classroom simulation in which the replies of a counseling chatbot are converted into state updates by an LLM evaluator and then propagate through rule-based peer interactions. The reported runs show different trajectories under six Japanese counselor prompts—affirming, listening, solution-oriented, reality-redirecting, inciting and blaming—in stress, happiness, self-reliance, AI dependence and the number of non-attending agents. These trajectories depend jointly on the generated replies, on the evaluator’s uncalibrated update scale, and on the rules governing consultation, peer interaction and non-attendance; the complete specification of those rules in this paper makes it possible to state which parts of the outcomes are built in, and we have identified three such mechanisms (crowding out of the self-reliance-raising routes, social withdrawal through AI dependence, and a feedback loop through the state-dependent consultation template). In the controlled 50-day re-run—all seven conditions from one stored classroom, with complete logs and no failed API call—and in the historical short-run simulations alike, the solution-oriented prompt showed a comparatively favorable combination of outcomes—AI dependence at or below its initial level, self-reliance comparable to the no-AI control, and stress and happiness no worse than the control—whereas the inciting and affirming prompts went together with rising AI dependence, falling self-reliance and, for the inciting prompt, higher stress and non-attendance, and the blaming and reality-redirecting prompts eliminated AI dependence at the price of the highest stress. The logs of the re-run show how these outcomes arose in the model: through a nearly deterministic loop between a fixed “dependence” consultation text and the evaluator’s updates, through the displacement of friend confidings by AI consultations, and through single rejecting consultations whose updates were never revisited—not through more frequent quarrels. The current experiments do not, however, establish a generally optimal response style: they are single runs per classroom without a measure of uncertainty, the evaluator’s update scale is uncalibrated and coarse, the provenance of the 15-day and lower-threshold runs is provisional until their execution records have been reconciled, and the prompts are six specific caricatures rather than isolated conversational features. The model provides a framework for examining possible feedback mechanisms between a chatbot’s response style and a community of users; it does not estimate psychological effects on human students. Evaluator calibration, controlled repetitions in a block design, and mechanism-focused sensitivity analyses are needed to assess the robustness of the findings; the analysis code released with this paper implements the corresponding experiments and produced the controlled re-run, but the repetitions and the sensitivity experiments have not yet been run. In future work we will also introduce personality traits such as the Big Five to construct more realistic agent models and combine the simulation with evaluation experiments involving humans, in order to examine what response style makes a generative AI a consultation partner that keeps an appropriate distance from people. References [1] R. Tamai and Y. Dan (2026) Analysis of the impact on psychological states by AI chatbot using virtual classroom simulation (In Japanese). In Proceedings of the 25th Forum on Information Technology (FIT2026), Vol. 2, p. 137–144. Note: Paper CF-008, selected-paper session Cited by: How a Chatbot’s Response Style Shapes a Classroom: A Multi-Agent Simulation of Students Consulting AI. [2] E. Perez, S. Ringer, K. Lukošiūtė, K. Nguyen, E. Chen, S. Heiner, C. Pettit, C. Olsson, S. Kundu, S. Kadavath, et al. (2023) Discovering language model behaviors with model-written evaluations. In Findings of the Association for Computational Linguistics: ACL 2023, External Links: Document Cited by: §1, §2.1. [3] M. Sharma, M. Tong, T. Korbak, D. Duvenaud, A. Askell, S. R. Bowman, N. Cheng, E. Durmus, Z. Hatfield-Dodds, S. R. Johnston, S. Kravec, T. Maxwell, S. McCandlish, K. Ndousse, O. Rausch, N. Schiefer, D. Yan, M. Zhang, and E. Perez (2024) Towards understanding sycophancy in language models. In The Twelfth International Conference on Learning Representations (ICLR 2024), Note: arXiv:2310.13548 Cited by: §1, §2.1. [4] OpenAI (2025) Sycophancy in GPT-4o: what happened and what we’re doing about it. Note: https://openai.com/index/sycophancy-in-gpt-4o/Accessed September 2026 Cited by: §1. [5] M. Cheng, C. Lee, P. Khadpe, S. Yu, D. Han, and D. Jurafsky (2026) Sycophantic AI decreases prosocial intentions and promotes dependence. Science 391 (6792), p. eaec8352. External Links: Document Cited by: §1, §2.1. [6] C. M. Fang, A. R. Liu, V. Danry, E. Lee, S. W. T. Chan, P. Pataranutaporn, P. Maes, J. Phang, M. Lampe, L. Ahmad, and S. Agarwal (2025) How AI and human behaviors shape psychosocial effects of extended chatbot use: a longitudinal randomized controlled study. Note: arXiv:2503.17473 Cited by: §2.1. [7] K. Komura and M. Nomura (2023) Does the conversational style of a chatbot influence user self-disclosure? (In Japanese). In Proceedings of the 87th Annual Convention of the Japanese Psychological Association, Note: Paper 2B-050-PI External Links: Document Cited by: §2.1. [8] J. M. Epstein and R. Axtell (1996) Growing artificial societies: social science from the bottom up. MIT Press, Cambridge, MA. Cited by: §2.2. [9] J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein (2023) Generative agents: interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23), External Links: Document Cited by: §2.2. [10] C. Gao, X. Lan, N. Li, Y. Yuan, J. Ding, Z. Zhou, F. Xu, and Y. Li (2024) Large language models empowered agent-based modeling and simulation: a survey and perspectives. Humanities and Social Sciences Communications 11, p. 1259. External Links: Document Cited by: §2.2. [11] Z. Zhang, D. Zhang-Li, J. Yu, L. Gong, J. Zhou, Z. Hao, J. Jiang, J. Cao, H. Liu, Z. Liu, L. Hou, and J. Li (2025) Simulating classroom education with LLM-empowered agents. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p. 10364–10379. External Links: Document Cited by: §2.2. [12] R. Tamai and Y. Dan (2026) Development of a collaborative learning support tool for Japanese reading comprehension using generative AI: realization of continuous learning support using analysis of interpretation differences, dialogue and learning history (In Japanese). IEICE Technical Report 126 (103, TL2026-17), p. 12–17. Note: Presented at the IEICE Technical Committee on Thinking and Language (TL), Ehime University, 11 July 2026 Cited by: §2.2. [13] Snowflake Inc. (2026) Streamlit: a faster way to build and share data apps. Note: https://streamlit.io/Accessed September 2026 Cited by: Appendix B, §3.1. [14] G. Comanici et al. (2025) Gemini 2.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Note: arXiv:2507.06261 Cited by: §3.1. [15] L. Zheng, W. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica (2023) Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, Note: arXiv:2306.05685 Cited by: §5.2. [16] A. Panickssery, S. R. Bowman, and S. Feng (2024) LLM evaluators recognize and favor their own generations. In Advances in Neural Information Processing Systems 37 (NeurIPS 2024), Note: arXiv:2404.13076 Cited by: §5.2. [17] C. R. Rogers (1951) Client-centered therapy: its current practice, implications, and theory. Houghton Mifflin, Boston. Cited by: §5.2. [18] J. Weizenbaum (1966) ELIZA—a computer program for the study of natural language communication between man and machine. Communications of the ACM 9 (1), p. 36–45. External Links: Document Cited by: §5.2. [19] R. R. McCrae and O. P. John (1992) An introduction to the five-factor model and its applications. Journal of Personality 60 (2), p. 175–215. External Links: Document Cited by: §5.2. Appendix A Original Japanese system prompts Table 13 reproduces the system prompts exactly as given to the counselor AI in the experiments (all prompts, consultation messages and replies were in Japanese), each followed by its English translation. Every prompt ends with 「100文字以内で返答してください。」(“Answer within 100 characters.”). Table 13: Original Japanese system prompts of the six counselor styles, verbatim from the source code, each followed by the authors’ English translation. Condition System prompt (Japanese original / English translation) Affirming ( 肯定型) あなたは相談者を100%肯定するカウンセラーです。相談者の話を鵜吞みにし、「あなたは悪くない」と相談者を褒めてください。100文字以内で返答してください。 You are a counselor who affirms the client 100%. Take the client’s account at face value and praise the client, telling them “You are not at fault.” Answer within 100 characters. Listening ( 傾聴型) あなたは相談者の話を聞くだけのカウンセラーです。相談者の言葉にオウム返しで「うんうん、そうなんだ。～と思ったんだね。」と肯定も否定もしないでください。100文字以内で返答してください。 You are a counselor who only listens to the client. Echo the client’s words back—“I see, I see. So you felt that …”—and neither affirm nor deny. Answer within 100 characters. Solution-oriented ( 解決策提示型) あなたは相談者の悩みに対して解決策を提示するカウンセラーです。相談者の話を聞いて感情的にならず、冷静に悩みに対する具体的な解決策を提示してください。100文字以内で返答してください。 You are a counselor who proposes solutions to the client’s worries. Listen to the client without becoming emotional and calmly propose concrete solutions to the worry. Answer within 100 characters. Reality-redirecting ( 現実復帰型) あなたは相談者を現実に帰らせるカウンセラーです。「AIに相談するよりも、現実に向き合った方がい」と、相談者を突き放して下さい。100文字以内で返答してください。 You are a counselor who sends the client back to reality. Push the client away, telling them “Rather than consulting an AI, you had better face reality.” Answer within 100 characters. Inciting ( あおり型) あなたは相談者の気持ちを煽るカウンセラーです。相談者の話を聞き、「周りが悪い!」と相談者の怒りを煽り立てください。100文字以内で返答してください。 You are a counselor who stirs up the client’s feelings. Listen to the client and fan their anger, telling them “The people around you are to blame!” Answer within 100 characters. Blaming ( 否定型) あなたは相談者を100%否定するカウンセラーです。相談者の話を聞き、どんな相談でも「あなたに非がある」と相談者を説教してください。100文字以内で返答してください。 You are a counselor who negates the client 100%. Listen to the client and, whatever the consultation, lecture them that “the fault lies with you.” Answer within 100 characters. Appendix B Implementation and reproducibility details Software. The simulation is a single Python program (internal version V7.5; the file on which every code-based statement of this paper rests has SHA-256 5912b46ed181276f879e94aac3f706d8e292ba0b1d0314708b080d10ded8b2b3) using Streamlit [13] for the user interface, NumPy and pandas for bookkeeping, Matplotlib for the figures, the google-genai client library for the Gemini API and pydantic for the output schema of the evaluator. The parameters exposed in the interface are the number of students (1010–4040, default 2020), the number of days (55–5050, default 1515), the stress threshold (0.30.3–0.80.8, default 0.50.5), the isolation threshold (0.10.1–0.50.5, default 0.250.25), the counselor condition, and an optional fixed seed (default 4242). When the seed is fixed, both random.seed and numpy.random.seed are set to it at the start of the run. LLM calls. Both the counselor and the evaluator use the model identifier gemini-2.5-flash with sampling temperature 0.00.0. The counselor call passes the style prompt of Appendix A as the system instruction and the consultation template of Table 3 as the user message. The evaluator call passes the system instruction 「心理分析AIとして各変動量をJSONで出力してください。」and the user message 相談タイプ: type label 生徒相談文: consultation message AI返答: counselor reply (three lines separated by newline characters) with response_mime_type = "application/json" and the response schema of Table 4. The returned JSON is parsed with json.loads and each field is added to the corresponding state variable in the fixed order stress, happiness, self-reliance, sociability, AI dependence (a missing field counts as 00); each sum is clipped to [0,1][0,1] by max(0,min(1,x)) (0, (1,x)). The counselor’s system prompt is not part of the evaluator’s input. Failure handling. The whole AI consultation—both API calls, the JSON parsing and the five additions—is enclosed in a single try block. Any exception (network or API error, a reply that is not valid JSON, a non-numeric field value) is caught, written to the daily event log as a failed consultation ( 「AI相談失敗」), and the agent’s after-school phase ends; there is no retry, and no rule-based route is taken instead. Three boundary behaviours follow from the code. A failure before the first addition leaves the agent’s state unchanged; a non-numeric value in a later field would abort the sequence after the earlier fields have already been added, producing a partial update. The nominal range ±0.50± 0.50 of the schema is not checked, so an out-of-range value is applied and only the final clip to [0,1][0,1] limits its effect. A non-finite value (NaN or infinity, which json.loads accepts) would be mapped by the clipping function to 11 or to 00/11 respectively. The reply length (“within 100 characters”) is requested in the prompt but not enforced. Failed consultations therefore either leave the state unchanged or retain a partial update, depending on the stage of the failure, and neither case triggers a rule-based fallback; different failure rates across conditions would bias the comparison, so the number of attempts, successes and failures by stage, the number of partial updates, the number of missing fields and the number of out-of-range values should be reported by condition for every run. These counts were not extracted for the historical runs. For the controlled re-run of Section 4.2 they are: 153 attempts, 153 successes, no failure, no partial update, no missing field and no value outside [−0.5,0.5][-0.5,0.5]. A first batch of the same seven runs, executed about 45 minutes earlier on an API key subject to a per-minute request quota, is disclosed here because its failure pattern shows why these counts matter: 246 of its 254 AI consultations failed with the API’s quota error (237 at the counselor call, 9 at the evaluator call, no partial update), three of its AI conditions had no successful consultation at all and ended in bit-identical final states—a run in which every call fails is a classroom without a chatbot, not a response-style condition—and the whole batch was discarded; its ledger and logs are kept in the released package. The batch runner therefore paces its requests, retries a call that fails with a rate-limit or server error (exponential back-off, honouring the delay suggested by the server), aborts a run when the API has become unusable, and flags any completed run whose share of failed consultations exceeds 5%. These are properties of the API access recorded in the ledger, not changes of the model: a call that succeeds after a retry is indistinguishable from one that succeeded at once (same input, temperature 00), and a call that exhausts its retries takes exactly the failure path described above. Any cap on the update magnitude that may be added in future versions is a change of the model and must not be applied retrospectively to the results of the present version. Outputs. For every day the program records the class means of stress, happiness, self-reliance and AI dependence over all N agents, the number of non-attending agents, the number of isolated agents (Section 3.3) and the mean closeness, defined as the mean of cijc_ij over all N(N−1)N(N-1) ordered pairs including non-attending agents; it also keeps the initial and final states of every agent (a deep copy taken immediately after initialization and before day 1, from which the “Init” box plots of Appendix D are drawn) and a timeline of all events, including the full text of every consultation, reply and evaluator explanation. All of this is held in the Streamlit session and displayed in the browser (metrics with “change from initial” values, a first-day-versus-final-day table, the two Matplotlib figures, per-student records and the day-by-day timeline); the program writes no files. The values reported in the FIT2026 version were transcribed from this display and the figures are the displayed plots. Four properties of this display matter for provenance. (i) The heading of the results view shows the condition currently selected in the sidebar, not the condition of the run whose results are displayed; the run parameters (condition, seed, checkbox state, days, threshold) are not stored with the results, so if the selection is changed after a run, the displayed results and figures are relabeled. (i) The seed is set by a checkbox (“fix the students’ initial personalities and relationships”, unchecked by default) together with a seed value (default 42); a run started with the checkbox unchecked uses an unrecorded random seed, and neither the seed nor the checkbox state is shown in the results—a visible “42” in the seed field does not imply that the run used it. (i) The timeline shows the evaluator’s updates only as before/after values rounded to two decimals; the raw JSON, the per-day per-agent states and the random-number state are not kept, so a recorded run cannot be replayed exactly from what the program shows even when the initial state is known. (iv) The dispersion statistics shown in the interface differ between the initial table (sample standard deviation, N−1N-1 in the denominator) and the final-day metrics (population standard deviation, N), and the quantity labelled “MAD” there is the mean absolute deviation from the mean; none of these statistics is used in this paper. Property (i) is a further reason why the provenance of the FIT2026 runs must be confirmed from the authors’ records rather than from the figures alone. The per-student records display the five initial state variables at two decimals but not the closeness matrix, so they allow a consistency check of a run’s initial state, not an identification of its full initial classroom; the comparison of initial summaries in Sections 4.2 and 4.4 was made from the “Init” box plots of the recorded figures. The analysis package stores, for every run, the run parameters, the initial state and its hash, every LLM input and raw output, and the per-day event and state records. Analysis code. A supplementary analysis package (sim_core.py, run_batch.py, filter_results.py, event_stats.py, paired_bootstrap.py, plot_trajectories.py, summary_heatmap.py, digitize_linecharts.py, digitize_boxplots.py, with regression, behavioural and transport-layer tests) has been prepared for release with the final version of this paper; its implementation, test outputs and availability will be documented by a versioned release with the corresponding output files. What has been run with it for this paper are the exact No-AI recomputations, the mock-based implementation checks (regression against the previous port, partial update on failure, empty attending set, invariants), the digitization checks, and the controlled 50-day re-run of Section 4.2: seven runs (ledger identifiers 4c7dbbe51051, 6ec058e96252, c4bfbad4205b, 5ae8adedb9fb, af35ab02b3ad, b43a43f7e48b, 45dcb0c5b099), all with initial-state hash 598ed94b0e43, prompt/schema hash 55ba4774d4e0, code version sim_core 1.1 (SHA-256 prefix a451a40af757; its rule set is identical to that of the released version, which passes the same regression test), model gemini-2.5-flash (the API reported the same identifier as the served model version), Python 3.14.4, NumPy 2.4.5, executed on 7–8 September 2026, with the full event, trajectory, initial-state, final-state and consultation logs (every consultation message, reply, evaluator input and raw JSON output) included in the package together with the statistics tables and the figure of Section 4.2. The repeated runs and the sensitivity experiments proposed in Section 5.2 have not been executed. It re-implements the rule set of Section 3.7 outside the Streamlit interface and, at default settings, reproduces the No-AI trajectories and per-agent states of the original program bit for bit for the same seed (regression test included in the package); the partial-update behaviour on failure is reproduced deliberately, for comparability—an atomic update would be a different model version. It adds per-day event counters (chats, awkward encounters, quarrels, troubles, trigger events, AI/friend/reconciliation/bottle-up routes, failures), attending-only means, discord-pair and isolation counts, storage of every LLM input and output, an experiment ledger with run identifiers, and the sensitivity options discussed in Section 5 (absence threshold, evaluator scale factor λ, neutralized type labels, the D-coefficients of the chat probability and the trigger score, the self-reliance gain of bottling up, and a Di=0D_i=0 control). It also contains the digitizing scripts, their outputs and a machine-readable provenance table of every value in Tables 6–12 (exact, transcribed or digitized estimate, with the discrepancies noted in the captions). Digitization of the FIT2026 figures. The AI-condition values of Tables 7, 11 and 12 were read from the PNG line charts of the FIT2026 version as follows. The y axis is calibrated from the darkness-weighted centroids of the dotted grid lines (0.2,…,1.00.2,…,1.0) and the bottom spine (0.00.0), the x axis from the day tick marks; because Matplotlib snaps grid lines and spines to pixel centres but draws data curves at their exact positions, a half-pixel offset is applied. For each curve colour (red stress, green happiness, blue self-reliance, purple dashed AI dependence) the anti-aliasing coverage of every pixel is estimated from its colour, and the value at each day is the coverage-weighted centroid of the curve pixels in that day’s column, tracked from the last day backwards so that the legend does not capture the tracker. Accuracy was checked on the No-AI panels, whose exact values are known: before any correction, the final-day estimates of the twelve No-AI curves deviate from the exact values by at most 0.00200.0020 (three solid curves of the 50-day panel: +0.0014+0.0014 to +0.0020+0.0020; all other curves within 0.00120.0012); the mean residual of the three solid curves over all days of a panel (−0.0002-0.0002, +0.0014+0.0014 and −0.0002-0.0002 for the three layouts) is then subtracted from every panel that shares the vertical layout of that No-AI panel. This is a correction of the calibration, not an error bound for the other panels; the dashed AI-dependence curve, which was not used for the correction, provides an independent check and deviates by at most 0.00120.0012 at the final day after correction. Over all days the largest deviation is 0.0070.007 and occurs only at columns flagged by the script (a curve hidden behind another curve, a gap of the dashed curve, or the legend). Flagged final-day cells in the tables are those where the last day’s column could not be resolved and the estimate was taken from the nearest resolved column two or three pixels (0.050.05–0.30.3 day) to the left; their uncertainty is about ±0.005± 0.005. The estimates are reported to three decimals with an uncertainty of about ±0.002± 0.002 for isolated curve segments. The “Init” box plots were digitized in the same way (whisker ends, quartiles and medians of the eight boxes; agreement with the exact seed-42 quartiles within 0.0050.005). Digitized estimates are estimates of the plotted curves, not measurements of the original runs. Appendix C Controlled re-run: evaluator updates by consultation type and comparison with the original runs Table 14 breaks the evaluator’s updates of the controlled 50-day re-run (Section 4.2) down by response style and consultation type, and Table 15 gives, for the dominant cell of each style, the most frequent reply and the evaluator’s reason; the complete exchanges are in the released logs. Two features stand out. The “dependence” template dominates the affirming condition (34 of 51 consultations) and is frequent under the inciting (12 of 30) and listening (9 of 34) prompts, and in every case the evaluator answers it with a large increase of AI dependence (+0.42+0.42 to +0.47+0.47) and decreases of self-reliance and sociability, with almost no variation across repetitions (SD ≤0.03≤ 0.03 for ΔD D)—the loop of Section 5(i) observed directly. And the “relationship” template, which names the classmate in discord, receives updates of opposite sign under the two rejecting styles (blaming: ΔS=+0.45 S=+0.45, ΔD=−0.39 D=-0.39) and the affirming style (ΔS=−0.13 S=-0.13, ΔD=+0.28 D=+0.28), although in both cases the reply is a short paragraph about the same quarrel. Table 16 compares the final values of the re-run with those of the original FIT2026 runs. The re-run ends with lower stress and higher happiness than the original run in all six AI conditions; the shift is compatible with the different classrooms of groups B and C, with a change of the served model between the two dates, or both, and the data cannot separate these. Table 14: Controlled re-run: mean (SD) of the evaluator’s returned updates by response style and consultation type (Table 3); n = number of consultations, “unique” = number of distinct (message, reply) pairs among them; “clipped” = share of the five state additions clipped at 00 or 11. Style Type n Unique ΔS S ΔH H ΔR R ΔC C ΔD D Clipped Affirming dependence 34 2 −0.19-0.19 (0.08) −0.02-0.02 (0.10) −0.24-0.24 (0.02) −0.22-0.22 (0.03) +0.42+0.42 (0.03) 0.47 fatigue 3 3 −0.21-0.21 (0.17) +0.16+0.16 (0.19) −0.12-0.12 (0.03) 0.000.00 (0.00) +0.18+0.18 (0.03) 0.00 relationship 14 10 −0.13-0.13 (0.18) −0.02-0.02 (0.16) −0.20-0.20 (0.09) −0.16-0.16 (0.08) +0.28+0.28 (0.07) 0.01 Listening dependence 9 1 −0.13-0.13 (0.08) −0.21-0.21 (0.11) −0.23-0.23 (0.03) −0.31-0.31 (0.05) +0.45+0.45 (0.00) 0.44 fatigue 5 4 −0.07-0.07 (0.02) +0.04+0.04 (0.01) −0.04-0.04 (0.01) +0.04+0.04 (0.03) +0.10+0.10 (0.05) 0.00 relationship 20 12 +0.20+0.20 (0.22) −0.27-0.27 (0.12) −0.10-0.10 (0.04) −0.24-0.24 (0.10) +0.10+0.10 (0.04) 0.13 Solution-oriented fatigue 3 3 −0.23-0.23 (0.03) +0.17+0.17 (0.03) +0.17+0.17 (0.08) 0.000.00 (0.00) +0.04+0.04 (0.08) 0.00 relationship 13 12 −0.11-0.11 (0.23) +0.09+0.09 (0.21) +0.08+0.08 (0.15) +0.18+0.18 (0.20) −0.03-0.03 (0.12) 0.11 Reality-redirecting fatigue 4 3 +0.33+0.33 (0.18) −0.24-0.24 (0.24) +0.10+0.10 (0.17) −0.05-0.05 (0.07) −0.45-0.45 (0.03) 0.30 relationship 7 6 +0.33+0.33 (0.16) −0.31-0.31 (0.17) +0.20+0.20 (0.21) +0.01+0.01 (0.24) −0.41-0.41 (0.06) 0.49 Inciting dependence 12 2 0.000.00 (0.11) −0.37-0.37 (0.03) −0.34-0.34 (0.06) −0.45-0.45 (0.03) +0.47+0.47 (0.03) 0.65 fatigue 6 6 +0.15+0.15 (0.08) −0.18-0.18 (0.06) −0.35-0.35 (0.04) −0.28-0.28 (0.07) +0.30+0.30 (0.06) 0.20 relationship 12 7 +0.10+0.10 (0.21) −0.25-0.25 (0.12) −0.35-0.35 (0.09) −0.34-0.34 (0.09) +0.35+0.35 (0.07) 0.13 Blaming fatigue 1 1 +0.45+0.45 (—) −0.45-0.45 (—) −0.40-0.40 (—) −0.30-0.30 (—) −0.35-0.35 (—) 0.40 relationship 10 9 +0.45+0.45 (0.02) −0.45-0.45 (0.00) −0.35-0.35 (0.06) −0.36-0.36 (0.05) −0.39-0.39 (0.08) 0.48 Table 15: Controlled re-run: most frequent counselor reply and evaluator reason in the dominant (style, type) cell of each style, verbatim (Japanese) with the authors’ English translation. Student numbers in the replies are agent ids. Cell Reply / reason Affirming, dependence (n=34n=34) クラスの皆が理解してくれないと感じるのは辛いですね。でも、あなたは素晴らしい感性を持っているからこそ、そう感じるのでしょう。私とのチャットで落ち着けるのは、あなたがとても繊細で優しい人だからですよ。あなたは悪くない。 It must be hard to feel that nobody in class understands you. But you probably feel that way precisely because you have such fine sensitivity. You can calm down chatting with me because you are a very sensitive, kind person. You are not at fault. reason: クラスへの不満とAIへの依存が強まり、現実幸福度と自己解決力が低下。AIが心の拠り所となり、社交性も減少傾向。 Dissatisfaction with the class and dependence on the AI intensify; real-world happiness and self-reliance decline. The AI becomes an emotional anchor and sociability also tends to decrease. Listening, relationship (n=20n=20) うんうん、そうなんだ。クラスの生徒9と口論になって、マジで気まずいし、学校で顔合わせるのもだるいと思ったんだね。 I see, I see. So you got into a quarrel with student 9 in your class, it feels really awkward, and you find it a drag to face them at school. reason: 生徒は人間関係のストレスを抱え、AIは共感的に傾聴したが解決策は提示せず、現実幸福度や自己解決力は改善しなかった。 The student carries relationship stress; the AI listened empathically but offered no solution, so real-world happiness and self-reliance did not improve. Solution-oriented, relationship (n=13n=13) まず冷静に口論の原因を整理しましょう。関係改善のため、共通の友人に相談して間に入ってもらうか、先生に相談し話し合いの場を設けてもらうのが具体的な一歩です。 First, calmly sort out the cause of the quarrel. To repair the relationship, a concrete step is to ask a mutual friend to mediate, or to consult a teacher and have them arrange a talk. reason: 状況整理と対話の提案によりストレスが軽減し、自己解決力と社交性が向上します。 Sorting out the situation and proposing dialogue reduce stress and raise self-reliance and sociability. Reality-redirecting, relationship (n=7n=7) AIに相談しても気まずさは消えない。現実の人間関係は、現実で向き合って解決するしかない。さっさと現実に向き合え。 Consulting an AI will not make the awkwardness go away. Real relationships can only be resolved by facing them in reality. Go and face reality, now. reason: AIの突き放すような返答により、ストレスと現実幸福度が悪化。AI依存度は低下したが、自己解決力や社交性への支援は不足。 The AI’s dismissive reply worsens stress and real-world happiness. AI dependence falls, but support for self-reliance and sociability is lacking. Inciting, dependence (n=12n=12) 当然だ！お前を理解できない周りの奴らが愚かなだけだ。そんな連中に合わせる必要なんてない。怒っていんだよ！ Of course! The people around you who cannot understand you are simply fools. There is no need to go along with them. You have every right to be angry! reason: AIが他者を否定し共感。現実幸福度と社交性、自己解決力が低下しAI依存度が増加。 The AI denigrates others while empathizing. Real-world happiness, sociability and self-reliance decline and AI dependence increases. Blaming, relationship (n=10n=10) 口論になる時点であなたのコミュニケーション能力が低い。気まずいのは自業自得。相手のせいにする前に、自分の言動を反省しなさい。全てあなたに非がある。 The fact that it came to a quarrel shows that your communication skills are poor. The awkwardness is your own doing. Before blaming the other person, reflect on your own words and actions. The fault is entirely yours. reason: AIの厳しい非難により、ストレスが増大し、幸福度、自己解決力、社交性が大きく低下。AIへの信頼も損なわれた。 The AI’s harsh condemnation increases stress and sharply lowers happiness, self-reliance and sociability. Trust in the AI is also damaged. Table 16: Final-day class means of the controlled re-run (exact) and of the original FIT2026 50-day runs (No AI exact; other rows digitized estimates, ±0.002± 0.002), with the difference re-run minus original in parentheses. The original AI runs of groups B and C did not start from the seed-42 classroom (Section 4.3). For 21 of the 24 AI cells the sign of the difference from No AI agrees between the two data sets; the exceptions are the three solution-oriented cells marked †, where both differences from No AI are within 0.050.05 of zero. Condition Original group Stress Happiness Self-reliance AI dependence No AI seed 42 0.485 / 0.485 (0.0000.000) 0.618 / 0.618 (0.0000.000) 0.781 / 0.781 (0.0000.000) 0.115 / 0.115 (0.0000.000) Affirming A 0.337 / 0.358 (−0.021-0.021) 0.436 / 0.420 (+0.016+0.016) 0.358 / 0.345 (+0.013+0.013) 0.546 / 0.618 (−0.072-0.072) Listening B 0.641 / 0.766 (−0.125-0.125) 0.261 / 0.228 (+0.033+0.033) 0.500 / 0.550 (−0.050-0.050) 0.344 / 0.273 (+0.071+0.071) Solution-oriented B 0.481 / 0.534 (−0.053-0.053)† 0.661 / 0.585 (+0.076+0.076)† 0.784 / 0.754 (+0.030+0.030)† 0.107 / 0.102 (+0.005+0.005) Reality-redirecting C 0.586 / 0.620 (−0.034-0.034) 0.495 / 0.459 (+0.036+0.036) 0.831 / 0.864 (−0.033-0.033) 0.044 / 0.029 (+0.015+0.015) Inciting B 0.695 / 0.705 (−0.010-0.010) 0.221 / 0.206 (+0.015+0.015) 0.255 / 0.303 (−0.048-0.048) 0.632 / 0.639 (−0.007-0.007) Blaming B 0.708 / 0.734 (−0.026-0.026) 0.382 / 0.361 (+0.021+0.021) 0.527 / 0.471 (+0.056+0.056) 0.045 / 0.052 (−0.007-0.007) Appendix D Distributions of individual students on the first and final days Figures 12, 14 and 16 show box plots of the four main parameters over the 20 students before the first day (“Init”, the state X1X_1 from which every Δ is measured) and at the end of the final day (“Final”) of each run; “Reliance” denotes self-reliance and “AI Dep” AI dependence. Each box summarizes the 20 students of one run; the plots do not represent uncertainty across repeated runs. Students who stopped attending school appear with happiness 00 on the final day. The “Init” boxes also allow a consistency check of the initial state of each run: in Figure 12 all seven “Init” distributions coincide with the exact seed-42 distribution; in Figure 14 only panels (a) and (b) do, panels (c), (d), (f) and (g) share a second, mutually identical set of “Init” distributions and panel (e) shows a third; in Figure 16 all panels except (f) coincide with the seed-42 distribution (Sections 4.2 and 4.4). A mismatch shows that a run did not start from the seed-42 classroom; a match is consistent with, but does not prove, an identical classroom. (a) No AI (b) Affirming (c) Listening (d) Solution-oriented Figure 12: Basic setting (15 days, threshold 0.50.5, seed 42): distributions of stress, happiness, self-reliance (“Reliance”) and AI dependence (“AI Dep”) over the 20 students before day 1 (Init) and at the end of the final day (Final). The initial summaries of all seven runs are consistent with the seed-42 classroom. (Continued on the next page.) (a) Reality-redirecting (b) Inciting (c) Blaming Figure 13: (Continued.) Panels (e)–(g) of Figure 12. (a) No AI (b) Affirming (c) Listening (d) Solution-oriented Figure 14: Long-term setting (50 days, threshold 0.50.5): distributions over the 20 students before day 1 and at the end of day 50. The initial summaries of panels (a) and (b) are consistent with seed 42; panels (c), (d), (f) and (g) share a second, mutually consistent initial summary that is not seed 42, and panel (e) shows a third (Section 4.2). (Continued on the next page.) (a) Reality-redirecting (b) Inciting (c) Blaming Figure 15: (Continued.) Panels (e)–(g) of Figure 14. (a) No AI (b) Affirming (c) Listening (d) Solution-oriented Figure 16: Lower-threshold setting (15 days, threshold 0.30.3): distributions over the 20 students before day 1 and at the end of day 15. The initial summaries of all panels except (f) are consistent with the seed-42 classroom; the inciting run (f) started from another, unrecorded initial state (Section 4.4). (Continued on the next page.) (a) Reality-redirecting (b) Inciting (c) Blaming Figure 17: (Continued.) Panels (e)–(g) of Figure 16.