What changed
Researchers tested a custom tutor called PS2 Pal against an established active-learning class in which students worked in small groups and received feedback from instructors and course staff. The study ran during two class meetings in an introductory physics course for life-science students. One group used the tutor at home for a lesson on surface tension while the other learned in class; the next week, for fluid flow, the groups switched.
The worksheets and underlying content were the same in both conditions. On immediate tests scored out of 5, the AI condition had a median post-test score of 4.5, compared with 3.5 after the in-class condition. Against a combined pre-test median of 2.75, the paper reports that the median gain with the tutor was more than twice the in-class gain. Analyses adjusting for prior physics performance, topic, test version, ChatGPT experience, and time on task also favored the tutor.
What this could change for you
For students, the practical idea is a self-paced first encounter with difficult material. The median tutor session lasted 49 minutes; researchers treated the in-class learning period as 60 minutes after subtracting test time. Seventy percent of tutor sessions took less than 60 minutes, while students who needed longer could keep working.
For teachers and course designers, the result points to a specific use: prepare students for class, then reserve shared time for discussion, projects, synthesis, and difficult questions. It does not say that opening a general chatbot will reproduce the result. The team supplied verified step-by-step solutions, controlled the activity sequence, embedded instructional videos, and prompted the model to manage cognitive load and encourage active effort.
What it does not prove
This was one course at one highly selective university, covering two introductory physics topics over two weeks. The researcher-written tests measured learning immediately after each lesson; the trial did not measure long-term retention, performance across a semester, other subjects, collaboration skills, or results in younger students.
The comparison cannot isolate which ingredient caused the advantage. The AI condition differed in pace, location, delivery medium, feedback, videos, and individual versus group work. Eligibility required consent, participation in both formats, and completion of every pre- and post-test, leaving 194 of 233 enrolled students in the study population.
The bottom line
A carefully constrained AI tutor beat a strong classroom comparison on immediate tests for two physics lessons, and 70% of tutor sessions took less than the assumed 60-minute class learning period. That supports structured, teacher-built tutoring as a promising way to introduce material before class—not generic chatbots, durable learning gains, or replacing teachers.
Primary research
AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
Scientific Reports · 2025 · DOI 10.1038/s41598-025-97652-6
