Make Meaningful Feedback Possible at Scale with AI

This study tested whether an AI-generated feedback message targeting a student's specific wrong answer helps middle school math students more than simply telling them their answer was incorrect. The study was conducted in the Fall and Winter of the 2025-26 school year and included 21,478 students and over 100,000 problem-solving attempts in which students submitted an answer and received feedback.

‍

The researchers compared how students who received the message and those who did not answered the next problem on the first try, without requesting a hint or explanation. They divided students into groups based on recent scores on the same skill. Students with intensive learning needs improved significantly after receiving feedback, more than those who did not. Students who were mostly on track did not improve significantly on the next problem, but their performance did not decrease either. 

‍

AI-written feedback performed as well as teacher-written feedback, with no difference in student performance between the two. What differed was cost. We estimate it would cost $1.42 per message to hire a person compared to $0.05 for the AI. In other words it costs 28 times as much to pay teachers. Conclusion: Targeted AI-generated feedback helped the students who were struggling most, without getting in the way of stronger students, at a fraction of the cost.

‍

‍