How assigned AI use before class shapes active student engagement in class

Synthesis: Wang and colleagues report a preregistered field experiment — the first large-scale randomized trial of its kind, they argue — in which 759 MBA students across ten sections of a required strategy course were each randomly assigned two of ten case discussions to prepare for with CAiSEY, a purpose-built voice-based AI discussion partner that argues the opposing position in a spoken exchange. Teaching assistants counted voluntary contributions live, excluding instructor cold calls. After a student's second assigned use, voluntary contributions in later sessions rose about 31% relative to the sample mean of 0.73 per session — but nothing changed in the session the tool prepared, and participation fell after a first use alone. Heavier users also reported greater comfort speaking up and greater perceived learning, but no change in focus or motivation. The work reframes AI in education around engagement as an intermediate learning outcome rather than Assessment-based test scores, tying the effect to lower anxiety and greater Self-Efficacy.

Key Findings

1. A second use, not a first, moved behavior. Contributions rose in the next session (β₂ = 0.183) and all future sessions (β₂ = 0.227, both p < 0.05) only after the second assigned opportunity; a first use preceded lower participation (β₁ = −0.151).
2. The effect is about 31%. The qualifying-use estimate for all future sessions corresponds to roughly one extra voluntary contribution every four to five sessions, 31% of the 0.73-contribution mean.
3. Nothing changed in the session CAiSEY prepared. Same-session estimates were near zero for assignment and qualifying use alike (β₁ = −0.040, β₂ = −0.036), so the benefit is lagged.
4. Take-up was high, but dose was uneven. Of the 504 students whose two assigned sessions were observed, 92.2%–98.9% submitted a conversation, but only 38.1%–67.0% held one of at least ten minutes.
5. Comfort and perceived learning rose; focus and motivation did not. Later qualifying use predicted greater comfort (λ = 0.131) and deeper perceived learning (λ = 0.124, both p < 0.05); focus (λ = −0.035) and motivation (λ = 0.040) stayed flat.
6. The result was robust to the definition of use. A second assigned opportunity predicted more participation at every conversation-length threshold from four to ten minutes (β₂ = 0.099 to 0.227, all p < 0.05).

The intervention: a voice partner that argues back

CAiSEY was purpose-built for the course rather than a general chatbot. Each case came with an opening question; the student chose a position and the tool adopted the opposing one, so the student defended a stance against an adversarial voice AI partner until they ended the exchange. The platform then generated a summary and the student submitted a written reflection. Instructors and teaching assistants were blind to who was assigned. The design targeted two barriers at once: mastery of the case material and the anxiety that keeps students from speaking, which prior work frames as fear of negative evaluation. The study thus randomizes when, not whether, each student practiced — a pedagogical agent used as a rehearsal partner rather than a tutor.

Why class participation is the outcome, not a test score

The authors argue AI-and-learning research has over-indexed on exam and homework scores, which say little about the engagement behaviors that mediate learning. Speaking up in class is both a performance to prepare for and a habit that generalizes to career settings where decisions are made orally, linking participation to professional training and employability. Prior reviews connect voluntary contributions to Self-Efficacy and Critical Thinking, while fear of negative evaluation suppresses them. That is why self-report items about comfort and perceived learning are read alongside the behavioral counts: they test a proposed mechanism, not just a feeling. The paper leans on active learning scholarship and the ICAP idea that explaining deepens understanding.

A lagged effect that needs a second exposure

The dose-response is the paper's least intuitive result. Preparing with CAiSEY did nothing for the case it was assigned to; a lone first use preceded lower voluntary participation (β₁ = −0.151 in the next session); only the second use predicted more contributions afterward. The estimate for all future sessions (β₂ = 0.227 under the randomization-instrument approach) is around 31% of the 0.73-contribution average. Because every student received two assignments, the comparison is between earlier and later assignment, which the authors read as use versus not-yet-use. The gain held at every threshold from four to ten minutes, so long conversations were not required.

Where this sits in the AI-and-learning debate

The study occupies a contested space. Much recent evidence, including work on generative AI in secondary mathematics, warns that unguarded tools invite offloading and can damage learning, and a Nature feature the authors cite worries about eroding critical thinking. This study points the other way, but for a tool built with deliberate Guardrails: purpose-built, adversarial, and tied to a specific discussion the student is about to have. The authors caution that the result does not generalize to generic chatbots used without design, and that Generative AI effects depend heavily on deployment. They position voice AI as provoking the productive friction critical thinking depends on rather than replacing it — supported here for oral rehearsal, still open for broader simulated dialogue.

What this means for practice

• Instructors. Build a pre-class rehearsal step into a discussion you will actually hold, and make it adversarial — the student picks a position and the tool argues the other side — rather than asking students to summarize the reading.
• Instructors. Budget for repetition: the gain appeared only after a second use and surfaced in later sessions, so a one-off AI assignment is unlikely to change who speaks.
• Course and program designers. Voice-based rehearsal is a scalable complement to timetabled speaking practice for large higher-education cohorts, supporting a live oral skill rather than substituting for it.
• Administrators. Because the effect persists across the rest of the course, treat pre-class AI preparation as an intervention with carryover and evaluate it against participation data your classes already generate.

Limitations

• Only two exposures, and no never-used comparison. Every student was assigned CAiSEY twice, so the study cannot compare end-of-term learning outcomes for users against students who never used the tool; the contrast is use versus not-yet-use.
• No human-partner control condition. The authors would have preferred to compare the AI partner with a human rehearsal partner, but standardizing and monitoring that was beyond their capacity.
• One purpose-built tool in one course. CAiSEY was tailored to the case-based discussions of a single required course at one US business school taught by four instructors, so results need not transfer to generic chatbots used without design.
• Participation was coded live by teaching assistants. Two sections (152 students) did not record cold calls separately, one retained section's attendance had to be reconstructed, and voluntary contributions were inferred by subtracting cold calls.
• The self-reported outcomes are perceptions. Comfort, focus, motivation and perceived learning came from seven-point self-report items, which the study treats as subjective experience rather than validated skill.
• The causal reading rests on assumptions. The local-average-treatment-effect interpretation requires an exclusion restriction, and reading later-use coefficients as end-of-term change assumes ratings would not otherwise have moved.

Citation

Wang, D. J., Modi Jain, N., Burbano, V., Guzman, J., Keum, D., Kim, S., Kogut, B., & Wright, N. (2026). How assigned AI use before class shapes active student engagement in class. arXiv preprint arXiv:2610.10463.