Imagine three students working through the same physics problem. One understands the physics but struggles with mathematics. Another can manipulate the equations but is unsure why they apply. The third has solved the problem and wants to explore what happens when an assumption changes. They are studying the same material, but they need different help.
A tutor sitting with each of them could respond to what that student needs, adjusting the next question or explanation as the conversation develops. Giving every student in a classroom that kind of attention is something educators have been working toward for decades.
In 1984, Benjamin Bloom asked whether teachers could bring the benefits of one-to-one tutoring to an entire class. He examined tutoring alongside mastery learning, in which teachers checked what students understood and gave further help where they found gaps. Both approaches gave students opportunities to work through material they had found difficult [1].
Computers offered another route. Researchers were already experimenting with computer tutors in the 1970s. Over the following decades, they explored different ways these systems could help students work through problems. Some checked a student’s final answer; others offered guidance during the steps leading to it [2]. Designing these tutors meant thinking through familiar teaching questions. When should a student get a hint? When would it help to see a worked example? Research on feedback examined how these forms of help could support learning [3].
By the 2010s, researchers were testing systems such as Cognitive Tutor and ASSISTments with thousands of students across dozens of schools. Computer-supported tutoring had grown from small demonstrations into a substantial field of research [4,5].
Generative AI has renewed interest in that longstanding ambition. A student might ask, “I understand how you rearranged the equation, but why does that equation apply here?” A responsive tutor could explain the principle, ask the student to apply it, and use the reply to decide what to discuss next. The promise is to make exchanges like this available to many more learners.
Three approaches to AI tutoring
Three recent experiments explore different ways of providing this support. Harvard researchers tested an AI-supported physics lesson [6]. A University of Pennsylvania team studied mathematics practice at a Turkish high school [7]. At Stanford, researchers built a tool to assist human tutors [8]. For each study, we ask how students were helped, what they did with that help, and what they could do afterward on their own.
Harvard: a lesson at the student’s pace
Greg Kestin, Kelly Miller, and colleagues tested their AI tutor in an introductory physics course for life-science students at Harvard. They focused on two lessons: surface tension and fluid flow. Their analysis included results from 194 students. The researchers kept students with their usual classroom study partners and randomly assigned these small groups to begin either in class or with the AI tutor at home. The following week, they switched formats for the second lesson. Students took a test before and immediately after each lesson [6].
In class, students worked through the problems together in small groups, with help from the teaching staff. After each activity, the instructor answered their questions and went over the parts they had found difficult. Students using the AI tutor studied the same material and used the same worksheet. They watched recorded introductions and worked through the problems with the tutor [6].
The researchers gave the tutor solutions prepared by the course instructors and designed the software to take students through each activity one part at a time. They instructed it to keep replies brief, respond to the student’s question, and explain one step at a time. One activity asked how an enlarged model of a microscopic organism would have to move to reproduce the organism’s behavior in fluid. The instructors supplied both the solution and the reasoning behind it so the tutor could help students work through the problem [6].
On the tests taken immediately after each lesson, the median, or middle score, was 4.5 out of 6 after AI tutoring and 3.5 after classroom instruction. The difference remained after the researchers accounted for students’ prior knowledge and other factors in their statistical analysis [6].
Students spent a median of 49 minutes with the AI tutor, compared with 60 minutes set aside for learning in class. The authors suggest that being able to work at their own pace and ask for help when they needed it may have contributed to the higher scores. Because the study tested these features together, it cannot tell us how much each mattered [6].
The higher scores are encouraging. Because the study covered two lessons and tested students immediately afterward, it leaves open whether the advantage would last across a full course or on a test taken weeks later [6].
University of Pennsylvania: from practice to independent work
What happens when a student finishes practicing with help and then faces a similar problem alone? Hamsa Bastani, Osbert Bastani, and colleagues took this question to a Turkish high school, where nearly 1,000 students took part over four 90-minute mathematics sessions [7].
Each class was assigned one of three ways to practice. Students in the control group used textbooks and notes. A second group could also ask GPT Base, a relatively unrestricted GPT-4 chatbot, for help. The third used GPT Tutor. The researchers gave this version the teachers’ worked solutions and guidance on common mistakes and instructed it to guide students with hints [7].
Each session began with the teacher reviewing material taught earlier in the course. Students then worked through practice problems using the materials and tools available to their group. The teacher briefly went over the correct answers before students took an exam on closely related problems, working without books, laptops, or outside help [7].
Both groups using AI scored higher during practice. After the researchers accounted for prior grades and other factors, students using GPT Base scored an estimated 13.7 percentage points higher than the control group, while those using GPT Tutor scored 36.1 points higher. On the exam without help, students who had used GPT Base scored an estimated 5.4 percentage points lower than the control group. Those who had used GPT Tutor scored 0.4 points lower, but that difference was not statistically significant [7].
All three groups received a teacher-led review of the answers before the exam. The experiment varied the support available during practice. When the researchers examined the conversations, they found signs that many students had used GPT Base to get solutions. They suggest that some students may have missed the practice they needed to solve similar problems on their own. GPT Tutor combined several forms of guidance, so the study cannot tell us how much each contributed [7].
The two assessments tell different parts of the story. Both tools helped students score higher during practice. On the independent exam, however, students who had used GPT Tutor showed no statistically detectable advantage over the control group [7]. This makes the practice itself worth a closer look. We would like to understand how students work through a difficult step with a tutor’s help, and how that experience can prepare them to approach a similar problem on their own.
Stanford: supporting the human tutor
Rose E. Wang, Ana T. Ribeiro, and colleagues at Stanford asked how AI could help human tutors decide what to say next. The November 2025 version of their working paper reports results from 783 tutors and 1,013 students. Tutors were randomly assigned to receive access to Tutor CoPilot or to teach without the tool [8].
The researchers partnered with FEV Tutor, an online service that connects students with human tutors. During a mathematics session, a tutor could ask the tool for suggestions based on the conversation so far: a guiding question, an explanation, a hint. The tutor could use a suggestion as written, edit it, or ignore it [8].
Near the end of a lesson, students were expected to take a short test, called an exit ticket, without their tutor’s help. In calculating the pass rate, the researchers counted unattempted tests as not passed. Giving tutors access to Tutor CoPilot increased the estimated pass rate by four percentage points, from about 62% to 66%. That is roughly four more tests passed for every hundred comparable tutoring sessions [8].
Tutors chose to use the tool in about 29% of the sessions where they had access to it. The estimate includes both the sessions in which they used it and those in which they did not [8].
The gain was larger, nine percentage points, for students working with lower-rated tutors. The conversations changed too: tutors with the tool more often asked students to explain their reasoning, and less often offered generic encouragement [8].
The researchers found no statistically significant improvement on the end-of-year mathematics tests. The study lasted about two months, and students could work with tutors from either group across different sessions. Students therefore had broadly similar exposure to tutors with AI access. The authors suggest that the short duration and limited differences in exposure may have made a longer-term effect difficult to detect [8].
These findings suggest that AI could help tutors think through what to ask or explain next. Longer studies could show whether the improvements last and whether using the tool helps tutors develop their own teaching skills.
What these studies suggest
The three studies show different ways AI can become part of a lesson. At Harvard, students worked through physics problems with an AI tutor. In the Pennsylvania study, students could use AI while they practiced mathematics. At Stanford, tutors received suggestions they could use in their conversations with students. Each study involved different learners and teaching circumstances, which matter when we interpret the results. Together, they offer several ways to explore how AI might help people learn [6–8].
Teachers helped shape the learning experience in all three studies. Instructors had prepared the solutions and guidance used by Harvard’s tutor and Pennsylvania’s GPT Tutor. In the Stanford study, tutors decided which suggestions to use and how to respond to the student [6–8]. Anyone developing this kind of support faces familiar teaching decisions: what to explain, when to offer help, and how to respond as a student begins to understand.
We need to be just as thoughtful about how we judge learning. During practice, a score tells us what students can do with the help available. A test taken independently shows what they can do without that help. Testing them several weeks later tells us what they remember. The Pennsylvania results show why these questions need to be considered separately [7]. In future studies, we would want to examine each of them and see whether students can use what they have learned in an unfamiliar problem.
Think again of the three students at the start. The first, after working through the mathematics with help, could try a similar calculation alone. The second could explain why the equation applies and name its assumptions. The third could change an assumption and work out how it affects the solution. A tutor would need to recognize these different needs and respond to each student.
We would also want to know what each student could still explain or solve several weeks later. Could they recognize an error in a convincing explanation? At Criticality Labs, we hope to explore these questions in nuclear engineering and the radiation sciences. Our larger ambition echoes Bloom’s question from 1984: to make individual attention widely available, and to personalize learning.
Sources
[1] Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16. doi: 10.3102/0013189X013006004
[2] VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. doi: 10.1080/00461520.2011.611369
[3] Shute, V. J. (2008). Focus on formative feedback. Review of Educational Research, 78(1), 153–189. doi: 10.3102/0034654307313795
[4] Pane, J. F., Griffin, B. A., McCaffrey, D. F., and Karam, R. (2014). Effectiveness of Cognitive Tutor Algebra I at scale. Educational Evaluation and Policy Analysis, 36(2), 127–144. doi: 10.3102/0162373713507480
[5] Roschelle, J., Feng, M., Murphy, R. F., and Mason, C. A. (2016). Online mathematics homework increases student achievement. AERA Open, 2(4), 1–12. doi: 10.1177/2332858416673968
[6] Kestin, G., Miller, K., Klales, A., Milbourne, T., and Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. Supplementary Material 1 also consulted. doi: 10.1038/s41598-025-97652-6
[7] Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi: 10.1073/pnas.2422633122
[8] Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., and Demszky, D. (2025). Tutor CoPilot: A human-AI approach for scaling real-time expertise. EdWorkingPaper, November 2025 revision. doi: 10.26300/81nh-8262