By Steve Grubbs with editing and research by ChatGPT

At Harvard University, one of the most rigorous tests of AI tutoring to date has reshaped how educators think about personalized learning and intelligent tutor systems. 

In a randomized controlled trial led by Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti, researchers set out to measure whether a custom-built AI tutor could match—or even outperform—the best active-learning classroom instruction. What they found wasn’t just impressive; it was paradigm-shifting.

The study took place in Harvard’s Physical Sciences 2 course, the school’s largest introductory physics class, during the Fall 2023 semester. Roughly 180 students alternated between two conditions each week: one in-class, guided by highly refined active-learning methods, and another at home, where they used a purpose-built AI tutor. The AI was powered by a large language model but carefully designed with expert-authored scaffolds, step-by-step reasoning, and built-in guardrails to avoid hallucination.

The result? Students using the AI tutor achieved roughly twice the learning gains of those in the active-learning classroom. On standardized pre- and post-tests covering surface tension and fluid flow, the median post-test score in the AI group rose from 2.75 to 4.5 points, while the in-class group improved from 2.75 to 3.5. Statistical analysis confirmed a large and highly significant advantage for AI instruction, with effect sizes approaching 0.7–1.3 standard deviations—levels of impact rarely seen in educational research.

Even more striking was the efficiency. Students using the AI tutor learned more while spending less time on task: a median of 49 minutes versus a 60-minute in-class block. The researchers found no correlation between time spent and post-test scores, suggesting that the AI’s strength lay in adaptive pacing and immediate feedback rather than longer study hours. In other words, it wasn’t quantity of time—it was quality of interaction. I was busy in law school and any time I could buy back was well worth it.

Student sentiment mirrored the data. On engagement and motivation scales, students rated the AI tutor higher than traditional in-class learning—4.1 versus 3.6 for engagement and 3.4 versus 3.1 for motivation. A remarkable 83 percent of students said the AI’s explanations were as good as or better than their instructors’ in class. That statistic alone challenges the long-standing assumption that meaningful learning requires a human always at the center of instruction.

What made this success possible wasn’t raw computing power; it was intentional instructional design. The Harvard team combined LLM intelligence with structured pedagogy—expert prompts, micro-scaffolded feedback, and short contextual videos—to model not just what students should learn, but how they learn best. By embedding proven active-learning methods within the AI framework, they effectively created a scalable digital tutor that both personalizes and democratizes high-quality instruction.

Note that this was not a question-prompting AI tutor, but it fully taught a lesson complete with demonstratives and a conversational 2-way instructor. Most of what we see in the marketplace today is text-based question prompting.

The implications extend far beyond one physics course. When an AI tutor can match or exceed one of the world’s most advanced classroom models—while adapting to every learner in real time—it signals a revolution in accessibility and effectiveness. Whether in rural schools, workforce training, or higher-education labs, this model points toward a future where every learner can have a tutor as skilled as Harvard’s best faculty, available anytime, anywhere.

For educators, technologists, and policymakers, the message is clear: the era of scalable, personalized AI tutoring has arrived—not as a novelty, but as a proven accelerator of human learning. 

To take an advanced AI tutor out for a test drive, try HoloTutor at https://holotutor.ai 

Citation:

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, Article 17458. https://doi.org/10.1038/s41598-025-97652-6

Steve Grubbs is CEO of VictoryXR, a global leader in immersive AI education with emphasis on VR simulations. With 25 years in education policy and ed-tech, he’s driven the company’s mission to bring virtual laboratories, simulations, and spatial learning to students worldwide.