The Personal Tutor at Scale: What Harvard’s AI Experiment Could Mean for the Future of Education

For generations, education has confronted a fundamental scaling problem. We know that individualized instruction can be extraordinarily effective, but a teacher responsible for twenty, thirty, or several hundred students cannot simultaneously become a personal tutor for every learner.

Artificial intelligence may be beginning to change that equation.

A randomized controlled trial conducted at Harvard University offers unusually compelling evidence that a carefully engineered generative-AI tutor can substantially improve learning under at least some conditions. More importantly, the research raises a much larger question: What happens if individualized AI tutoring can eventually be extended across subjects and throughout education—from elementary reading and mathematics through high school, college, and lifelong technical education?

The implications could extend far beyond improving test scores. At sufficient scale, personalized AI-assisted education could potentially contribute to a population better equipped to understand science and technology, adopt new technologies productively, evaluate their consequences, and manage their risks.

But the Harvard research also contains an important warning: simply giving students access to a general-purpose chatbot is not what produced the result.

The Harvard Experiment

The study is:

Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti, “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting,” Scientific Reports, Volume 15, Article 17458, published June 3, 2025.

Read the complete Harvard study at PubMed Central

DOI: 10.1038/s41598-025-97652-6.

The research involved students in Physical Sciences 2, Harvard’s largest introductory physics course. The investigators began with 233 enrolled students; 194 met the study’s inclusion requirements. The experiment was conducted during the fall 2023 semester and examined instruction involving two physics topics: surface tension and fluid flow. 

This distinction matters: the comparison was not between AI and a mediocre lecture course. Harvard’s control condition employed active learning, a well-established instructional methodology that itself generally produces better learning outcomes than traditional passive lectures. Students worked on activities in groups, received assistance from course staff, and received targeted instructor feedback. The researchers report that 89 percent of surveyed students considered the course more active-learning-oriented than other Harvard STEM courses they had taken.

The AI therefore faced a demanding benchmark.

How the Experiment Worked

The researchers used a randomized crossover design.

Students were divided into two groups. During one week, one group learned through the AI tutor while the other received the normal active-learning classroom instruction. The following week, the conditions were reversed. Consequently, students experienced both instructional approaches. 

Before each lesson, students completed a pre-test establishing their baseline knowledge. After instruction, they completed a post-test measuring mastery. They were also questioned about engagement, enjoyment, motivation, and growth mindset. 

The researchers attempted to keep the instructional content itself constant. The classroom and AI conditions used identical content and worksheets. Activity introductions were also identical in substance, although classroom students received them live while AI students received prerecorded versions. 

The fundamental variable was therefore the mechanism through which students interacted with the material and obtained guidance and feedback.

This Was Not Simply ChatGPT Teaching Physics

This is perhaps the most important technical detail in the entire experiment.

The researchers did not merely tell students to open a general-purpose chatbot and ask it to teach physics.

They constructed a specialized AI tutoring environment called PS2 Pal.

The system incorporated principles drawn from educational research, including active learning, management of cognitive load, growth-mindset encouragement, instructional scaffolding, timely feedback, and self-paced learning. 

The investigators discovered that prompt engineering alone was insufficient. A language model could sometimes discuss portions of a multipart problem in the wrong sequence or introduce material before the student was ready for it. The software therefore imposed additional structure, guiding students sequentially through the components of each activity. 

Accuracy was another engineering concern.

Rather than trusting GPT-4 to independently generate every physics solution, the researchers supplied comprehensive expert-developed, step-by-step answers that the AI could use in constructing explanations. This reduced dependence on unconstrained model reasoning and helped control the hallucination problem inherent in generative AI. 

This suggests an important principle for educational AI:

The educational system is not merely the model. It is the model plus curriculum, pedagogical architecture, validated knowledge, instructional sequencing, assessment, feedback, and appropriate controls.

That distinction becomes increasingly important if such systems are eventually deployed with children.

What Happened?

The results were substantial.

The combined pre-test median was 2.75. Students receiving classroom active learning achieved a median post-test score of 3.5, while students receiving AI tutoring achieved a median post-test score of 4.5.

Measured relative to the pre-test baseline, the median learning gain in the AI condition was more than twice that of the classroom active-learning condition. The difference was highly statistically significant, with p < 10^-8. 

Regression analysis controlling for prior physics knowledge, broader course performance, conceptual physics understanding, previous ChatGPT experience, lesson topic, test version, and time on task continued to show substantially better performance for the AI-tutored students. The researchers estimated an effect size ranging from approximately 0.73 to 1.3 standard deviations after accounting for ceiling effects.

And students generally did not need more time to obtain those gains.

The conventional lesson allocated approximately 60 minutes to learning after testing time was removed. The median AI-tutoring session lasted 49 minutes, and 70 percent of students using the AI tutor spent less than 60 minutes on the material.

The students also reported greater engagement and motivation in the AI condition. Engagement averaged 4.1 on a five-point scale for AI instruction compared with 3.6 for classroom active learning. Motivation averaged 3.4 versus 3.1. Differences in enjoyment and reported growth mindset were not statistically significant. 

This means the result was not simply greater measured learning accompanied by an obviously worse student experience.

Why Might Individualized AI Instruction Work?

One explanation is straightforward: classrooms operate synchronously, while individuals do not learn synchronously.

A teacher must select a pace.

That pace will inevitably be too fast for some students, appropriate for others, and unnecessarily slow for still others.

A personal tutor does not have that constraint.

Students who have already mastered prerequisite material can move rapidly. Students who encounter difficulty can stop. They can request another explanation. They can work through another example. They can revisit an earlier concept without delaying thirty classmates or feeling embarrassed about asking a question publicly.

The Harvard data contain an intriguing indication of precisely this effect. Students who considered the classroom pace too fast spent more than the median amount of time with the AI tutor, while those who considered the classroom pace too slow spent less than the AI median.

In effect, the instructional clock became individualized.

That may be one of AI tutoring’s most consequential characteristics.

What the Experiment Does Not Prove

The findings are impressive, but their boundaries are equally important.

This was not a multiyear experiment. It did not demonstrate long-term knowledge retention. It did not test elementary-school children learning to read. It did not demonstrate improved statewide standardized-test performance. It did not establish that AI tutoring will outperform excellent teachers across every subject.

The authors themselves explicitly caution against assuming that structured AI tutoring will always outperform classroom active learning, particularly for activities requiring synthesis across multiple concepts and higher-order critical thinking. 

Nor do they propose replacing teachers.

Their proposed model is closer to a technologically enhanced flipped classroom. AI can help students acquire and practice foundational knowledge individually, allowing teachers and classroom time to concentrate increasingly on advanced problem solving, projects, collaboration, critical thinking, discussion, synthesis, and other activities for which human interaction remains particularly valuable. 

That distinction should shape any attempt to extend the research.

Now Ask the Larger Question

Suppose this approach works beyond introductory college physics.

What happens if every elementary-school student has access to an individualized AI tutor working with the classroom teacher and curriculum?

Consider reading.

A classroom teacher may have students simultaneously struggling with phonemic awareness, decoding, vocabulary, fluency, comprehension, background knowledge, or combinations of these skills.

An appropriately designed educational system could continuously determine which skills an individual student has mastered, identify where difficulty is occurring, adjust the pace and complexity of instruction, provide additional practice, and measure whether an intervention is actually producing improvement.

The objective would not be to replace the teacher.

It would be to give every teacher something historically impossible to provide economically at population scale: a continuously available individualized instructional capability for every student.

The same possibility exists in mathematics.

If a fifth-grade student is struggling with fractions because of an earlier weakness in multiplication, simply repeating the current lesson may not solve the problem. A sufficiently capable tutoring system could potentially detect the prerequisite deficiency, move backward to repair it, verify mastery, and then return the student to grade-level material.

Conversely, a student who has already mastered the material need not remain artificially synchronized with the average classroom pace.

That produces a fundamentally different educational architecture.

Instead of:

Curriculum → Teacher → Classroom → Average Pace → Periodic Assessment

we could increasingly have:

Curriculum → Teacher + Individual AI Tutor → Continuous Student Interaction → Individualized Pace and Scaffolding → Continuous Formative Assessment → Identified Skill Gaps → Targeted Intervention → Verified Mastery → Advancement.

Could This Improve State Reading and Mathematics Proficiency?

That is now a testable research question.

The Harvard study does not establish that AI tutoring will raise elementary or secondary statewide proficiency scores. But it provides enough evidence to justify carefully designed longitudinal trials asking exactly that question.

Imagine following students across an academic year rather than two physics lessons.

Students could be assessed against established curriculum standards at the beginning of the year. The tutoring system could complement normal classroom instruction while maintaining a continuously updated representation of demonstrated skills and unresolved deficiencies.

Researchers could then measure not merely whether students liked the system, but whether it changed objective educational outcomes:

reading proficiency, mathematics proficiency, rate of mastery, retention, remediation success, attendance and engagement, teacher workload, achievement across different starting proficiency levels, and performance on independent standardized assessments.

The critical experiment would not ask whether students using AI can outperform students receiving poor instruction.

It should ask the much more demanding question Harvard asked:

Can carefully engineered personalized AI instruction improve outcomes when added to or compared with already effective educational practices?

From Elementary School to College

If the answer eventually proves to be yes across subjects and age groups, the implications become much larger.

The educational system could evolve toward a complementary division of labor.

Teachers would remain responsible for the human dimensions of education: leadership, motivation, judgment, social development, discussion, collaborative learning, mentorship, classroom culture, and sophisticated intellectual synthesis.

AI tutors could increasingly handle individualized repetition, explanation, practice, formative assessment, immediate feedback, prerequisite remediation, and self-paced progression.

This could continue through high school and university education.

And eventually it need not stop at graduation.

The same architecture could support lifelong education as workers encounter new technologies, occupations, scientific developments, and technical systems throughout their lives.

Toward a Technology-Mastering Civilization

This leads to the most consequential question raised by the Harvard experiment.

What happens when inexpensive individualized instruction becomes widely available throughout a person’s lifetime?

Civilizations increasingly depend upon technologies that relatively few citizens fully understand. Artificial intelligence, biotechnology, robotics, advanced manufacturing, cybersecurity, quantum technologies, autonomous systems, energy systems, and other emerging technologies are becoming simultaneously more powerful and more complex.

A society therefore faces two challenges.

It must be capable of creating and using increasingly sophisticated technology.

But it must also be capable of understanding and governing its consequences.

Those are not separate problems.

A technologically sophisticated population should be better positioned to understand what a technology can actually do, distinguish genuine hazards from imagined ones, recognize emerging risks earlier, evaluate evidence, participate intelligently in decisions about deployment, and develop technical controls that mitigate risks before significant harm occurs.

This suggests a concept worth considering: the technology-mastering civilization.

A technology-mastering civilization would not be a civilization that blindly embraces every new technology.

Nor would it be one that instinctively fears technologies because they are unfamiliar.

It would seek to develop sufficient scientific, mathematical, technical, historical, economic, and civic knowledge to understand new technologies, evaluate their potential benefits and consequences, use them productively, and establish proportionate safeguards against foreseeable accidental misuse, emergent failure, systemic risk, or intentional abuse.

Education is foundational to that capability.

The Potential Feedback Loop

If personalized AI tutoring ultimately improves educational performance, an important positive feedback mechanism becomes possible.

Better AI → Better Personalized Education → Greater Human Knowledge and Technical Capability → Better Innovation and Technology Management → Better AI and Other Technologies → Better Personalized Education.

That cycle could compound.

Better education could produce more scientists, engineers, technicians, entrepreneurs, researchers, teachers, and technologically capable citizens.

But its effect could extend beyond the technically trained population. AI tutoring could make sophisticated knowledge more accessible to people whose occupations are not primarily scientific or technical.

The objective would not be to turn everyone into an engineer.

It would be to raise society’s overall capacity to understand, learn, adapt, reason, and act competently in an increasingly technological environment.

Education as Civilization Infrastructure

For most of history, expert personal tutoring has been scarce because expert human attention is scarce.

Generative AI changes the economics of that constraint.

Once a well-designed tutoring capability exists, the marginal cost of providing another student with extensive individualized interaction can potentially become dramatically lower than providing that student with a dedicated human tutor.

That does not make teachers obsolete. It could make teachers more consequential.

A teacher equipped with individualized tutoring systems could potentially spend less time repeatedly delivering standardized information and more time performing functions requiring human expertise, judgment, mentorship, observation, motivation, collaboration, and intervention.

The teacher could become the leader of a much more capable instructional system.

And the student could cease being merely one member of a classroom moving at an averaged pace.

Each student could increasingly have an educational pathway responsive to what that student actually understands.

The Experiment We Should Now Run

The Harvard research should therefore be understood neither as proof that AI can replace education nor merely as another demonstration that large language models are impressive.

It demonstrates something narrower—and potentially much more important.

When generative AI was embedded within a carefully engineered pedagogical system, Harvard students learned substantially more in less time than students receiving an established active-learning classroom experience. 

That result deserves replication across institutions, subjects, student populations, age groups, and longer periods.

The next major research question is therefore not simply:

Can AI tutor students?

We increasingly know that under appropriate conditions it can.

The more important questions are:

Can individualized AI tutoring reliably improve educational performance throughout elementary, secondary, university, and lifelong education? Can it help teachers identify and correct learning deficiencies earlier? Can it raise reading, mathematics, science, and technical proficiency across entire populations? Can those gains persist over years rather than individual lessons?

And ultimately:

Could universally available individualized education help produce a civilization capable not merely of inventing increasingly powerful technologies, but of understanding them well enough to master their benefits, recognize their consequences, and control their risks?

The Harvard experiment cannot yet answer those questions.

But it provides a scientifically serious reason to start asking them.

And if the results ultimately scale beyond the Harvard classroom, the most consequential application of artificial intelligence may not simply be making machines more intelligent.

It may be using intelligent machines to help human beings become more capable learners.

References

1. Kestin, G., Miller, K., Klales, A., Milbourne, T., & 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. DOI: 10.1038/s41598-025-97652-6.
   Full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/
   Publisher: https://www.nature.com/articles/s41598-025-97652-6
2. Freeman, S., Eddy, S. L., McDonough, M., Smith, M. K., Okoroafor, N., Jordt, H., & Wenderoth, M. P. (2014). “Active learning increases student performance in science, engineering, and mathematics.” Proceedings of the National Academy of Sciences, 111(23), 8410–8415. DOI: 10.1073/pnas.1319030111.
   https://www.pnas.org/doi/10.1073/pnas.1319030111
3. 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.
   https://doi.org/10.3102/0013189X013006004
4. 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.
   https://doi.org/10.1080/00461520.2011.611369
5. Ma, W., Adesope, O. O., Nesbit, J. C., & Liu, Q. (2014). “Intelligent Tutoring Systems and Learning Outcomes: A Meta-Analysis.” Journal of Educational Psychology, 106(4), 901–918. DOI: 10.1037/a0037123.
   https://doi.org/10.1037/a0037123
6. Kulik, J. A., & Fletcher, J. D. (2016). “Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review.” Review of Educational Research, 86(1), 42–78. DOI: 10.3102/0034654315581420.
   https://doi.org/10.3102/0034654315581420
7. Black, P., & Wiliam, D. (1998). “Assessment and Classroom Learning.” Assessment in Education: Principles, Policy & Practice, 5(1), 7–74. DOI: 10.1080/0969595980050102.
   https://doi.org/10.1080/0969595980050102
8. National Reading Panel. (2000). Teaching Children to Read: An Evidence-Based Assessment of the Scientific Research Literature on Reading and Its Implications for Reading Instruction. National Institute of Child Health and Human Development, National Institutes of Health.
   https://www.nichd.nih.gov/publications/pubs/nrp/findings
9. National Mathematics Advisory Panel. (2008). Foundations for Success: The Final Report of the National Mathematics Advisory Panel. U.S. Department of Education.
   https://www2.ed.gov/about/bdscomm/list/mathpanel/report/final-report.pdf
10. U.S. Department of Education, Office of Educational Technology. (2023). Artificial Intelligence and the Future of Teaching and Learning: Insights and Recommendations. Washington, D.C.
    https://tech.ed.gov/ai-future-of-teaching-and-learning/
11. UNESCO. (2023). Guidance for Generative AI in Education and Research. United Nations Educational, Scientific and Cultural Organization, Paris.
    https://unesdoc.unesco.org/ark:/48223/pf0000386693
12. Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., et al. (2023). “ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education.” Learning and Individual Differences, 103, 102274. DOI: 10.1016/j.lindif.2023.102274.
    https://doi.org/10.1016/j.lindif.2023.102274
13. Lo, C. K. (2023). “What Is the Impact of ChatGPT on Education? A Rapid Review of the Literature.” Education Sciences, 13(4), 410. DOI: 10.3390/educsci13040410.
    https://www.mdpi.com/2227-7102/13/4/410
14. U.S. Department of Education, National Center for Education Statistics. National Assessment of Educational Progress (NAEP): The Nation’s Report Card. This national assessment provides longitudinal measures of U.S. student achievement, including reading and mathematics, and provides an appropriate measurement framework for investigating whether AI-supported personalized instruction ultimately produces population-level improvements in educational performance.
    https://www.nationsreportcard.gov/
15. National Academies of Sciences, Engineering, and Medicine. (2018). How People Learn II: Learners, Contexts, and Cultures. Washington, D.C.: The National Academies Press. DOI: 10.17226/24783. This synthesis of learning science provides broader evidence concerning prior knowledge, motivation, context, feedback, learning strategies, and the development of expertise relevant to the design of personalized educational systems.
    https://doi.org/10.17226/24783 

The first reference is the direct evidentiary foundation for the article; references 2–7 establish the research background on active learning, individualized tutoring, intelligent tutoring systems, and formative assessment; references 8–10 support the K–12 and educational-policy discussion; and the remaining references provide context for generative AI, measurement of educational outcomes, and learning science.