Intensive Tutoring Works, and 90% of American Students Don’t Have Access to It

A meta-analysis of 89 randomized controlled trials has just quantified what education practitioners have sensed for decades: high-dose tutoring produces an effect of +0.29 standard deviations on academic performance. It is one of the best-documented interventions in the entire education science literature. And yet, only one in ten American students benefits from it.

The obstacle is no longer scientific. It is political, financial, and organizational. We know what works. We know roughly how much it costs. We even know, by now, how to reduce this cost by a third through artificial intelligence tools without draining the program of its effectiveness. What is missing is the decision to do it at scale.

The Essentials

  • A meta-analysis of 89 randomized controlled trials (732 effect estimates, published between 1985 and 2020) measures an average effect of +0.29 standard deviations on academic outcomes for high-dose tutoring, signed by Andre Nickow, Philip Oreopoulos and Vincent Quan (NBER / Annenberg Institute at Brown University) and published in the American Educational Research Journal in 2024.
  • A randomized trial on 4,000 students shows that a hybrid model combining human tutors and AI retains 80% of this effect (+0.23 standard deviations in mathematics) for a cost reduced by one-third.
  • Only 10% of American students currently access structured tutoring, with a notable concentration in wealthy families in the private sector.
  • The main obstacle is no longer technical but political: who finances it, who is prioritized, and how to scale without diluting the effect.

+0.29 Standard Deviations: What This Number Really Changes

In education science, effects are rarely spectacular. Most well-evaluated interventions produce gains of 0.05 to 0.10 standard deviations. An effect of 0.20 is already considered substantial. At 0.29, high-dose tutoring places itself in a category apart.

To make it concrete: a gain of 0.29 standard deviations for a student at the 50th percentile moves them toward the 61st or 62nd percentile. Across an entire cohort, applied to the most behind students, it is the difference between dropping out and earning a diploma. For children who start with the least, every standard deviation gained counts twice.

The meta-analysis by Nickow, Oreopoulos and Quan (NBER / Annenberg Institute at Brown University), compiling 89 randomized controlled trials and 732 effect estimates between 1985 and 2020, provides two crucial insights beyond the average figure. First insight: the effect is robust. It holds across ages, subjects, socioeconomic contexts, and tutoring formats. Second insight: it is modulated by intensity. “High-dose” tutoring — generally defined as three or more sessions per week, often integrated into school time — systematically outperforms light formulas of one hour weekly.

This second point is essential. It explains why so many cheap, scattered, after-school tutoring programs do not reproduce the effects documented in the literature. The effect is not in the principle of tutoring. It is in its regularity and density.

Why It Works: The Mechanism Behind the Numbers

High-dose tutoring does not work because it would mysteriously be more effective than a good lecture. It works because it addresses a fundamental problem in mass education: the impossibility of adapting pace and content to each student when one teacher manages thirty simultaneously.

A tutor working with one or two students can identify the precise reasoning error that blocks a student’s understanding of fractions, correct it immediately, revisit it the next day if it persists, and adjust exercise difficulty in real time. This immediate, individualized feedback is what Benjamin Bloom called the “two sigma problem” in his 1980s work: students receiving individualized instruction outperform students in conventional classrooms by an average of two standard deviations. High-dose tutoring is the best approximation of this model at reasonable cost.

There is also a relational dimension that researchers have learned not to neglect. The programs that work best are those where the tutor is stable — the same adult each week, who knows the student, who has established some form of trust. Programs that underperform are those that rotate tutors, that treat the relationship as a secondary variable. This human dimension is what makes scaling difficult, and what risks being lost if costs are cut too quickly.

AI Reduces Cost by a Third Without Collapsing the Effect

The randomized trial conducted on 4,000 students, a collaboration between the University of Chicago Education Lab and several school districts, tests a direct hypothesis: can you hybridize human tutoring with AI tools to reduce cost without sacrificing effectiveness?

The result is encouraging and deserves to be read with precision. The effect measured in mathematics is +0.23 standard deviations in the hybrid model, versus +0.29 in the fully human model. This is a 21% reduction in effect for a 33% reduction in cost. The cost-effectiveness ratio improves. For a public decision-maker who must choose between reaching 100 students with full effect or 150 students with reduced effect, this is an arbitrage favorable to broad deployment.

The AI tool used in this trial does not replace the tutor. It assists: it prepares exercises calibrated to the student’s level, signals areas of persistent difficulty, manages the repetitive aspects of remediation so the human tutor can concentrate their time on moments of explanation and support. This is the separation between tasks that machines do better (consistency, patience, parametric adaptation) and tasks where humans remain irreplaceable (fine diagnosis, relationship, motivation).

This distinction matters, because poor policy decisions often consist of entirely replacing human tutors with adaptive software. Available data do not support this substitution. They support an augmentation. The reflection on what AI skills bring to workers is precisely of the same order: it is not about replacing, but augmenting what humans already do well.

10%: The Gap Between What We Know and What We Do

If high-dose tutoring is so effective, why do only 10% of American students have access to it? The short answer is: cost, the structure of the school system, and the absence of political will to change it.

Cost first. A quality high-dose tutoring program — three sessions per week, trained tutor, integrated into school time — costs between $1,500 and $4,000 per student per year depending on configuration. For a school district that spends an average of $13,000 per student, that is an additional cost of 10 to 30%. Reasonable for wealthy districts, prohibitive for poor ones — precisely those whose students would benefit most from the program.

Structure next. High-dose tutoring integrated into school time requires reorganizing schedules, training or recruiting tutors, coordinating with teachers. These are administrative frictions that school administrators, already under pressure, are reluctant to take on without resources and without clear mandate.

Political will finally. Since Covid-19, several American states have launched large-scale tutoring initiatives. Texas deployed its “Texas Learning Acceleration” program with federal funding, targeting several hundred thousand behind students. Louisiana, Tennessee, and Michigan have followed with varied approaches. These programs exist, they produce documented results, and they remain exceptions in a system that has not structurally decided that intensive tutoring was a standard, not an option.

Unequal access faithfully reproduces economic inequality. In the private sector and in wealthy families, individual tutoring has long been common practice: parents who can afford it buy it. What research says forcefully is that this advantage — long reserved for those who can pay — is reproducible at public scale for comparable unit cost. The market has solved the problem for 10% of students. Public policy could solve it for the remaining 90%.

Programs That Are Scaling Up

Several organizations have decided not to wait for the political debate to resolve itself. Their work is valuable because it documents not only effects, but the operational conditions for scaling.

The University of Chicago Education Lab has developed the “school-based tutoring” model which integrates tutoring into regular school schedules, on the model of a supplementary course rather than an after-school service. Their Chicago data shows substantial gains in mathematics among the most behind high school students, with effects maintained two years after the intervention.

High Dose Tutoring, a national organization, is working to standardize protocols and reduce implementation costs for mid-sized districts. Their approach involves pooling tutor training tools and curricula, so each district doesn’t have to reinvent the wheel.

AmeriCorps has included school tutoring in its deployment priorities since 2021, providing trained tutors at low cost in several hundred schools. This is an alternative funding model that partially circumvents the budgetary problem of districts: tutors are paid by the federal government, not by the school.

The Arnold Ventures foundation has been financing rigorous evaluations of these programs for several years, including the meta-analysis cited here. Its role is not insignificant: by financing costly RCTs that school districts cannot afford, it builds the scientific evidence that legitimizes public funding. This is the virtuous model between philanthropy and public policy — one that accelerates progress without substituting for democratic decision-making.

The question of large-scale funding remains open. The post-Covid American federal stimulus plan (ESSER) injected approximately $190 billion into school districts between 2020 and 2024, some of which funded tutoring programs. These funds are now exhausted. The question is whether states and districts will sustain these programs in their own budgets, or whether high-dose tutoring will return to its status as an exception reserved for pilot projects.

Industrialization That Preserves the Effect

There is a real risk in scaling, and it would be dishonest not to name it. When an educational program moves from pilot to industrialization, it tends to lose the conditions that explained its effectiveness. The best-trained tutors leave for well-funded pilot programs; replacements are less prepared. Supervision relaxes. Schedules lighten under union pressure or organizational constraints. Intensity drops, and with it, the effect.

This is not inevitable, but it is a documented trend in other areas of public policy, where best practices become diluted as they institutionalize. Think of what happened with apprenticeship in Germany, excellently designed in its original form and difficult to export without losing the institutional conditions that make it work.

Researchers at the University of Chicago Education Lab have identified the variables that predict effect preservation at large scale. Three stand out with consistency. The quality of initial tutor training and its regular renewal. Integration into school time rather than after-school, which guarantees frequency. And the stability of the tutor-student relationship for at least a full semester.

These three conditions have a cost. But they also have a measurable benefit: programs that respect them produce effects close to pilots; those that abandon them in the name of economy see their effects drop toward zero. The real risk is not failing to deploy tutoring at large scale. It is deploying it while giving up what makes it effective.

AI, correctly used, can help maintain these conditions. It can maintain exercise quality even when tutors are less experienced. It can flag students who are falling behind before the tutor detects it. It can reduce administrative time so the session is devoted to learning. These are real contributions, provided they are not confused with replacing human relationships.

What Decision-Makers Can Do Now

The blockage is not technological. The tools exist, the protocols are documented, the costs are known and alternative funding models (AmeriCorps, philanthropies, federal funds) are available. What is missing, in most American states, is an explicit decision to treat high-dose tutoring as basic educational infrastructure — the same as textbooks, teachers, or buildings.

Such a decision implies three concrete choices. First, sustainable funding inscribed in state budgets, not in one-off emergency funds. Next, explicit prioritization toward behind students in low-resource districts — those for whom the effect is greatest and access rarest. Finally, a results obligation anchored in rigorous evaluations, so funds do not serve to finance programs that abandon the conditions of effectiveness along the way.

This is not easy policy to implement in an extremely decentralized American education system, where school districts enjoy significant autonomy and federal funding is politically contested. But several states have shown it is feasible, and their data are beginning to convince their neighbors.

The real question, in the coming years, is not whether high-dose tutoring works. That question is settled. It is how many students of the next generation will grow up in a district that decided to make it accessible — and how many will grow up in a system that knew, and did not act.

Sources

  1. Nickow, Oreopoulos & Quan (2024) — meta-analysis of 89 RCTs on high-dose tutoring, American Educational Research Journal: https://journals.sagepub.com/doi/10.3102/00028312231208687
  2. University of Chicago Education Lab — RCT hybrid technology/tutor model (4,000 students): https://educationlab.uchicago.edu/resources/nber-working-paper-can-technology-facilitate-scale-evidence-from-a-randomized-evaluation-of-high-dosage-tutoring/
  3. University of Chicago Education Lab — school-based tutoring program, Chicago results: https://educationlab.uchicago.edu
  4. Benjamin Bloom — “The 2 Sigma Problem”, Educational Researcher, 1984: https://journals.sagepub.com/doi/10.3102/0013189X013006004
  5. Kraft (2020) — Interpreting Effect Sizes of Education Interventions: https://journals.sagepub.com/doi/10.3102/0013189X20912798
  6. NCES School Pulse Panel — 10% of American students benefiting from high-dose tutoring: https://www.nea.org/nea-today/all-news-articles/high-impact-tutoring
  7. AIBM — Article citing Nickow meta-analysis (Jan. 2026): https://aibm.org/research/the-strong-positive-effects-of-high-dose-tutoring-for-boys-and-girls/
  8. Texas Education Agency — Texas Learning Acceleration Grants
  9. AmeriCorps — national school tutoring program
  10. Arnold Ventures — funding RCT evaluations in education