AI-simulated students get stressed, happy, or hooked on chatbots — but does that match what real students go through?

This explores whether the stress, happiness, and AI-dependence shifts that LLM-agent students show in simulated classrooms match what real students go through when they use chatbots. The corpus has no direct test of this. Nothing here compares simulated trajectories against data from a real classroom. It does have the simulation itself, plus real-student findings and simulator-realism research that show what a fair comparison would need to check.

The simulation is a 20-agent classroom where six counselor-chatbot styles were run for 15 and 50 days. Each style produced a different path in stress, happiness, self-reliance, and AI dependence. The effects came from what the chatbot actually said, not from the style label, and they spread through peer interactions (How do different counselor styles shape student stress and AI dependence?). The chatbot side of that is the most trustworthy part, because those replies are real model output. The student side is the part that has to earn its realism. Real-world evidence suggests the mechanism is at least plausible. People reciprocate a chatbot's emotional disclosure the way they would a person's (Do chatbots trigger human reciprocity norms around self-disclosure?), so the social dynamics the simulation relies on are not made up.

The real-student evidence also sets a test the simulation hasn't been shown to pass. In one classroom study, students working with a chatbot did better on practical tasks and produced more knowledge-based dialogue than peer groups. They also said far less overall and expressed far fewer subjective views (Does chatbot interaction trade authenticity for better problem-solving?). If simulated students drift toward higher happiness or dependence without also going quieter and less personal, the simulation is missing something real. The corpus doesn't say whether it does.

There are three reasons to be cautious about long time spans and subtle states. First, real chatbot relationships lose their pull as novelty fades, and single-session findings don't extrapolate to medium or long use (Do chatbot relationships lose their appeal as novelty wears off?). Whether LLM agents get bored in the same way over 50 simulated days is an open question. Second, simulated users drift out of character over long conversations. Reinforcement-learning training cut that drift by 55%, which shows how large it was to begin with (Can training user simulators reduce persona drift in dialogue?). In a long run, that drift could look like a real "state change" when it is really an artifact. Third, LLMs are poor at recognizing ambivalence and early-stage motivation in real users (Why can't chatbots detect when users are ambivalent about change?). Models playing students may therefore flatten the hesitant, mixed-feeling responses that real classrooms are full of. That last point is my inference, not something the note tests.

The corpus does show how the gap could be closed. Conditioning a simulator on user-profile and intent variables, then checking realism with crowdsourced judges, discriminator models, and distribution matching, is a ready-made recipe (Can controlled latent variables make LLM user simulators realistic?). One reusable persona population can also be run through different formats, which allows like-for-like comparison (Can one persona population evaluate different application types?). An LLM scorer whose rubric ratings match human raters could measure simulated and real student interactions on the same scale (Can AI teammates assess collaboration without losing naturalness?). The simulation currently shows which factors could matter, chiefly the chatbot's actual wording and peer spread. It doesn't yet show how large the effects are, or how they play out over months.