How do different counselor styles shape student stress and AI dependence?

The paper treats a chatbot's response style as something that can be varied on purpose and followed through a community over time. It builds a virtual classroom of 20 student agents who chat, quarrel and consult friends by rule and, "when their stress is high," may consult a counselor AI (Gemini 2.5 Flash) instead. The counselor is given one of six style prompts: affirming, listening, solution-oriented, reality-redirecting, inciting or blaming. The conclusion says the runs "show different trajectories" under the six prompts in stress, happiness, self-reliance, AI dependence and the number of non-attending agents.

The mechanism is a two-stage pipeline. A second LLM call turns each exchange into updates of five state variables (stress, happiness, self-reliance, sociability and AI dependence) and never sees the style prompt, so the updates are driven by what the counselor said, not by its label. The updated states then "propagate through rule-based peer interactions," which is how one student's consultations can reach the rest of the class. Seven conditions, including a no-AI control, are compared over 15 days, over 50 days and under a lowered consultation threshold. The excerpt also lists a robustness protocol: ten independent classrooms, repeated LLM realizations of one classroom with its event stream held fixed, and evaluator updates scaled by 0.3 and 0.1.

The motivation is sycophancy. The introduction argues that assistants tuned for satisfaction may answer with "excessive empathy and affirmation that fosters dependence," and cites the April 2025 GPT-4o rollback. This extends Does chatbot interaction trade authenticity for better problem-solving?, which compared chatbot and peer conditions in real classrooms and could only ask whether the chatbot's passivity was the cause. Here the tone is the manipulated variable, and hostile and inciting styles serve as deliberate contrasts to the affirming one. It also gives a testbed for the frame-accepting stance described in How do chatbots enable distributed delusion differently than passive tools?: affirming and reality-redirecting are, in this framing, opposite ways of relating to a user's frame. The simulation resembles the multi-agent setup in Can psychotherapy actually teach AI chatbots better communication? but uses it differently. There it corrects the chatbot before it replies, and here it observes what the replies do to a population. AI dependence as an outcome that accumulates with repeated consultation also echoes Do humans learn to prefer AI partners over time?, which studied it with human participants.

The excerpt does not say which style raises or lowers any variable, how large the differences are, or whether they survive the robustness checks, because the results section is not included. It reports only that the trajectories differ. The state variables are LLM-evaluator judgments about simulated students, so any finding describes that simulation and not real classrooms. The sycophancy concern is stated as a motivation, not a result. At the strength the excerpt supports, the contribution is a design for treating conversational tone as an experimental variable with population-level consequences, and it is not evidence that affirmation harms.