The Big Picture
A chatbot that gives concrete, solution-focused replies lowered simulated students' dependence on the AI and increased self-reliance; overly affirming or emotionally amplifying replies increased dependence and, in some cases, stress.
ON THIS PAGE
The Evidence
Running a virtual classroom of simulated students, different chatbot reply styles produced clearly different classroom outcomes. Solution-oriented replies raised self-reliance and kept AI dependence low, while emotionally affirming or inciting replies drove higher AI dependence and sometimes higher stress and non-attendance. Calm, problem-focused responses tended to be best for long-term student resilience in the simulation. Semantic Capability Matching Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Mean stress in 15-day baseline runs ended between 0.40 and 0.43 (on the simulation's 0–1 scale).
2Self-reliance grew by about 0.04–0.07 in the 15-day runs via rule-based social routes.
3Agents with AI dependence ≥ 0.40 used a 'dependence' message that often fed a loop pushing dependence further, potentially toward 1.0.
What This Means
Engineers and product leads building conversational agents should care because reply style directly affects user dependence and self-reliance in this simulated setting. Researchers studying human-AI interaction and educators exploring AI tools for students can use these results to prioritize problem-solving assistance over excessive affirmation. Role-Based Agent Pattern
Key Figures
Fig 5: (a) No AI
Fig 6: (a) No AI
Fig 7: (a) No AI
Fig 8: (a) No AI
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Results come from a fully specified simulation that mixes rule-based student behavior with two calls to large language models for counseling and evaluation, so outcomes reflect modelled dynamics rather than measured human behavior. The evaluator that scores each exchange is itself an AI and has not been validated against human judgments, which may bias the measured state changes. Individual differences (personality, home environment) were not modeled, so effects may change with more realistic student variation. Orchestrator-Worker Pattern Mutual Validation Trap
Methodology & More
A virtual classroom was built where rule-based student agents interact with each other and sometimes consult a chatbot. Each consultation is answered by a counselor model under one of six stylized reply modes (affirming, listening, solution-focused, reality-redirecting, inciting, blaming) and then scored by a second model that updates each student's internal states: stress, happiness, self-reliance and AI dependence. Three experiment patterns were run: a 15-day baseline, a 50-day long run, and a high-frequency consultation scenario. LLM-as-Judge Pattern Across runs, solution-focused replies best supported simulated resilience: they limited increases in AI dependence while improving self-reliance and maintaining happiness. Emotion-focused styles that mainly affirmed or amplified student feelings increased AI dependence and, under the most extreme 'inciting' style, raised stress and non-attendance. The simulation also revealed built-in mechanisms that amplify dependence (e.g., dependent agents send a template that signals calmness from AI, which the evaluator then interprets as further dependence), so how often students consult the AI and how the AI phrases responses both matter. Because the evaluation itself is done by an AI and individual differences were not modeled, these are hypotheses worth testing with human studies before applying them in real classrooms. Mutual Verification Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
ArXiv preprint, no affiliations or citations provided and authors are not widely recognized in my knowledge — limited information.