When AI Tutors Speak: Evidence from a Randomized Field Experiment

In plain English

A field experiment with 86 MBA students pulled apart two things that usually travel together: how structured the AI tutoring was, and whether the student spoke or typed. Structure did the work — tutored students scored 6.63 points higher (p=.007), with relational-quality written answers rising from 8% to 49% versus 8% to 27% in the control group. Voice nearly doubled conversational engagement and cost 2.8x more to deliver, yet produced weekly mastery statistically equivalent to text. The practitioner takeaway is blunt: making an agent more humanlike does not by itself make it more educational, so the case for a voice or avatar layer has to rest on engagement, reach or accessibility rather than on learning gains.