Students spoke up more in later classes after a second chance to rehearse a case with a voice-based AI debate partner, according to an arXiv preprint posted Oct. 7. The field experiment assigned 759 first-year MBA students two opportunities to prepare for class discussions with the tool.
The researchers tracked whether students chose to speak during live class discussion. Before each assigned session, a student picked a position on a business case. CAiSEY, a purpose-built voice AI, took the opposing side, challenged the student's arguments and raised alternatives. After the spoken exchange, students wrote a reflection and received a summary and feedback.
Students were assigned two of ten case discussions in a required strategy course at Columbia Business School during fall 2025. The timing varied randomly, but everyone received the same number of opportunities. Instructors and teaching assistants did not know who had been assigned to use the tool. The participation analysis covered 4,787 class-session records from 504 students whose two assigned sessions were observed. Researchers counted voluntary comments and excluded instructor cold calls; the average was 0.73 comments per student per session.
The study found no increase in comments during the class students had prepared for, and no lasting gain after the first assigned opportunity alone. Across all later classes, the second assigned opportunity added an estimated 0.112 voluntary comments per student per session beyond the first opportunity, about 15% of the 0.73 average.
Among students whose assignment led them to use CAiSEY for at least ten minutes, a separate analysis estimated 0.227 additional comments per later session, about 31% of the average. That estimate applies to this assignment-responsive group, not every student.
In surveys completed by 627 students, later use was associated with greater reported comfort speaking up and perceived learning, but not with focus or motivation. Those measures were self-reported; the study did not establish that the tool improved grades or test scores.
The finding comes from one U.S. MBA strategy course and a discussion tool tailored to its cases. Because every student received two assigned chances to use CAiSEY, the researchers could not compare end-of-term learning outcomes with a group that never had access. They also did not compare AI practice with a human discussion partner. The paper is an arXiv preprint, and its results do not establish that a general-purpose voice chatbot would have the same effect.
Sources
- Dan J. Wang et al., “How assigned AI use before class shapes active student engagement in class” (arXiv, Oct. 7, 2026)
- Full paper (PDF)
Image credit: Oregon Department of Transportation via Wikimedia Commons, CC BY 2.0. Illustrative classroom photo; not the study participants.
.jpg){kind=link}