What One-to-One Tutoring Research Means for AI at Home

Research on one-to-one tutoring gives a real reason to take AI at home seriously, but no study shows that children who learn at home with AI end up "five years ahead." What the evidence supports is narrower: a well-designed tutor helps, and a tool that hands over answers can leave a student worse off.

This article goes through the main studies, in plain terms, with what each one measured and where it stops. Everything below was read on its publisher's page or in the paper itself on October 8, 2026.

Where does the idea of a huge tutoring advantage come from?

Most of it traces to a 1984 paper by Benjamin Bloom in Educational Researcher, "The 2 Sigma Problem." Bloom described studies in which students were randomly assigned to a conventional class of about 30 students, a mastery learning class, or tutoring with one tutor for each student (or for two or three students at once).

The average tutored student scored about two standard deviations above the average of the conventional class, which Bloom put above 98% of the students in that class. The mastery learning class, with the same class size but frequent feedback and corrective work, scored about one standard deviation above.

Two details are easy to miss. The studies used students in grades four, five and eight, in two subjects (Probability and Cartography), and each one limited instruction to 11 periods over a 3-week block. And Bloom did not present tutoring as a plan for everyone. He called the one-to-one version too costly to run at scale, and posed the "2 sigma problem" as a challenge to find cheaper methods that come close.

Do later studies of human tutoring find the same size of effect?

Not as large. A 2020 systematic review and meta-analysis of tutoring experiments for preK-12 students by Nickow, Oreopoulos and Quan found a pooled effect of 0.37 standard deviations. That is a solid, useful gain, and well under two.

The same review found that teacher and paraprofessional tutors had stronger effects than nonprofessional or parent tutors, that effects were generally largest in earlier grades, and that tutoring during the school day tended to beat after-school programs. In other words, who tutors and how it is set up changes the result.

Keep readingHow to Use AI for Learning at Home: A Complete Guide for ParentsThe best way to use AI for learning at home is as a helper the parent directs: it can draft plans, explain ideas, make practice questions and give feedback, while the parent decides what is taught an…

What have studies of AI tutors actually found?

Three published or preprint studies are the most cited so far. Each measured something specific, in a specific group.

A Harvard physics course

In a randomized crossover trial published in Scientific Reports on June 3, 2025, 194 eligible students in a Harvard introductory physics course did one lesson with a custom AI tutor and one with in-class active learning. The authors report that students learned significantly more in less time with the AI tutor: median time was 49 minutes against 60 minutes assumed for the class lesson. They estimate the effect at 0.63 standard deviations by one method and 0.73 to 1.3 by another.

The authors also list limits. It was one course at one elite university, the outcome was short-term, retention was not measured, and the tutor ran on expert-written, question-specific prompts that took months to build. They say AI may not have the same edge on tasks needing complex synthesis and higher-order critical thinking, and they advise against using it to fully replace in-class instruction.

A six-week program in Nigeria

A World Bank blog post from January 9, 2025 describes a randomized evaluation of a six-week after-school program in Edo, Nigeria, run in June and July 2024. Generative AI served as a virtual tutor, and the authors say it worked when implemented thoughtfully with teacher support. They report an effect of about 0.3 standard deviations, which they describe as roughly two years of typical learning.

The same post says the full results were still to be published, gives no sample size, and lists unknowns: long-term effects, effects in other subjects, and any negative or unintended effects. It also names teachers' role in the AI sessions as an open question.

Human tutors helped by AI

A preprint called Tutor CoPilot (version dated January 26, 2025) reports a randomized trial in live online tutoring where the AI coached human tutors rather than teaching children directly. Students whose tutors used it were 4 percentage points more likely to master topics, and 9 points for students of lower-rated tutors. The authors also report that tutor interviews flagged problems, such as suggestions that were not grade-level appropriate.

Can AI make learning worse?

Yes, under some conditions. A randomized controlled trial in a Turkish high school, published in PNAS in 2025, put nearly 1,000 students in grades 9 through 11 into three groups for four 90-minute math sessions: no AI, a ChatGPT-style GPT-4 interface ("GPT Base"), and a version prompted to give hints instead of answers ("GPT Tutor").

• During practice: GPT Base students scored about 48% better than the no-AI group, and GPT Tutor students about 127% better.
• On the closed-book exam: GPT Base students scored about 17% worse than the no-AI group, a statistically significant drop. GPT Tutor students were statistically indistinguishable from the no-AI group.
• Self-assessment: Students overestimated how much they had learned, especially in the GPT Tutor group.

The authors attribute the drop mainly to students using the open tool as a crutch for copying answers, not to mistakes by the model. They also note limits: one subject, one school, one country, short-term outcomes, and data from Fall 2023, when generative AI was new.

What does this mean for children learning at home?

None of the studies above followed homeschooled children, so any conclusion for a home setting is reasoning from these results, not a finding. The reasoning that holds up best from the stated facts:

• Design matters more than the brand name. The tutors that did well were built around hints, step-by-step solutions and set lesson goals. The open chat tool lowered exam scores in the Turkish trial.
• An adult stayed in the loop. The Nigerian program had teacher support, the Harvard lessons were designed by instructors, and Tutor CoPilot assisted human tutors. A parent who wants similar conditions may prefer to set the task and check understanding without the tool.
• Check learning without the tool. The PNAS gap appeared only when the AI was taken away. A parent who wants to know what a child has learned may prefer unaided questions to practice scores.
• Younger children are largely unstudied here. The AI tutor trials read for this article involved secondary and college students, plus a K-12 program in which AI supported human tutors.

Keep readingYour First Week Using AI for Learning at Home: A Day-by-Day PlanThe best first week with AI is a small one: one assistant, one subject, and a parent in the room. Use the plan below as a starting point and change the pace to fit your child. This plan is general ad…

How many children learn at home?

The National Center for Education Statistics estimated that 2.8% of students ages 5 to 17, about 1,457,000 children, were homeschooled in 2019, compared with 1.7% in 1999 and 3.4% in 2012. NCES notes that its 2019 figure leaves out students enrolled in school more than 24 hours a week and that Household Pulse Survey figures from 2020-21 are experimental and not comparable.