White paper | July 2026

Human Tutoring in an AI-Driven Learning Landscape

AI will change the economics and workflows of academic assistance, but it will not replace the relational, diagnostic, and accountable work of effective human tutoring. The durable model is the AI-augmented tutor.

Executive Summary

Artificial intelligence is rapidly changing how students seek help. Students use generative AI to search for information, brainstorm, edit writing, summarize readings, check homework, and obtain quick explanations. That shift is material for education. It does not, however, mean that human tutoring is becoming obsolete. The evidence points to a different conclusion: AI is likely to commoditize low-quality homework help while raising the value of human tutors who diagnose needs, build motivation, coach metacognition, and hold students accountable for durable learning.

This white paper advances three connected claims. First, human tutors will not be replaced by AI at the core of effective tutoring; they will be augmented by AI. The strongest emerging models position AI as a preparation, practice, feedback, and analytics layer under human judgment. Second, the students using AI in academics are not the same population, in the same moment, making the same decision that historically drove demand for paid human tutoring. AI use is broad, low-cost, and often task-specific. Paid tutoring has historically been narrower, more relationship-based, and more concentrated among families seeking sustained intervention, enrichment, test preparation, or accountability. Third, tutoring and teaching are distinct professional roles. This paper addresses tutoring exclusively: individualized or small-group academic support that supplements classroom instruction rather than replacing teachers or school systems.

The quantitative evidence is consistent with continued demand for human tutoring. Tutoring has one of the strongest evidence bases in education: a meta-analysis of field experiments found an average effect of 0.288 standard deviations across 89 randomized studies, with similar effects in literacy and mathematics (Nickow, Oreopoulos, & Quan, 2024). Earlier reviews likewise found that tutored students outperformed control students on academic outcomes and attitudes (Cohen, Kulik, & Kulik, 1982). Public schools have also continued to expand tutoring after pandemic learning disruptions: by May 2024, 87% of U.S. public schools reported offering some tutoring, and high-dosage tutoring availability rose from 39% to 46% during the 2023-24 school year (Institute of Education Sciences, 2024). Commercial market estimates show similar momentum, with global online tutoring services estimated at $10.42 billion in 2024 and projected to reach $23.73 billion by 2030 (Grand View Research, 2025a).

At the same time, AI adoption is becoming nearly mainstream among students. Pew Research Center found that 54% of U.S. teens reported using AI chatbots for schoolwork, while College Board reported that high school student use of generative AI rose from 79% to 84% between January and May 2025 (College Board, 2025; McClain et al., 2026). In the United Kingdom, the Higher Education Policy Institute reported that 92% of surveyed undergraduates used AI in some form in 2025, and 88% had used generative AI for assessments (Freeman, 2025). These figures show that AI is becoming a general academic utility. They do not show that AI is substituting one-for-one for human tutoring.

The future of tutoring is therefore not a contest between humans and machines. It is a design question: which tasks should be delegated to AI, and which responsibilities must remain human? AI can generate practice sets, translate explanations, surface patterns in student work, and provide always-on rehearsal. Human tutors are needed to interpret those outputs, correct errors, manage overreliance, build trust, recognize nonverbal and emotional cues, align support to classroom expectations, and help students transfer assisted performance into independent competence. The winning model is the AI-augmented tutor: a professional who uses technology to increase precision and reach while preserving the relational and cognitive work that makes tutoring powerful.

Evidence at a Glance

The table below summarizes the evidence base behind the paper's central conclusion: AI is reshaping academic assistance, but the strongest case is for augmentation of tutoring rather than replacement.

  
      Evidence area
      Quantitative signal
      Implication for tutoring
    
      Learning outcomes
      Tutoring meta-analysis: 89 randomized field studies; average effect = 0.288 SD. Earlier meta-analysis: 65 evaluations with positive academic and attitudinal effects.
      Human tutoring remains one of the strongest evidence-based interventions, especially when structured and frequent.
    
      School adoption
      87% of U.S. public schools offered tutoring in 2023-24; high-dosage tutoring rose from 39% to 46% during the year.
      Schools continue to invest in tutoring as a recovery and acceleration strategy.
    
      Market demand
      Global online tutoring estimated at $10.42B in 2024 and projected to $23.73B by 2030; U.S. online private tutoring projected from $4.33B to $8.09B.
      The market is growing despite the rise of generative AI.
    
      Student AI adoption
      54% of U.S. teens used chatbots for schoolwork; 84% of high school students reported generative AI use by May 2025; 92% of surveyed U.K. undergraduates used AI in some form.
      AI is becoming a broad academic utility, not simply a tutoring substitute.
    
      Market segmentation
      Only about 6-7% of U.S. families with children ages 6-17 paid for tutoring in 2022; tutoring centers were concentrated in higher-income areas.
      The historic tutoring population and the AI-use population are different in scale, access, and jobs-to-be-done.

Sources: Nickow et al. (2024); Cohen et al. (1982); Institute of Education Sciences (2024); Grand View Research (2025a, 2025b); College Board (2025); Freeman (2025); Kim et al. (2024); McClain et al. (2026).

Scope: Tutoring Is Not Teaching

Any credible discussion of AI and education must separate tutoring from teaching. The two roles overlap in their concern for student learning, but they differ in purpose, context, accountability, and unit of work. A teacher is responsible for a classroom or course: designing and sequencing instruction, managing a learning community, assessing performance against standards, communicating with families, adapting to institutional requirements, and sustaining a yearlong or semester-long arc of learning. A tutor, by contrast, works in a narrower and more personalized zone. Tutoring typically supplements classroom instruction through one-on-one or small-group support aimed at specific learning goals, skill gaps, assignments, or confidence barriers (National Student Support Accelerator, n.d.).

This distinction matters because claims about AI replacing teachers often bundle together tasks that have different risk profiles. AI can draft lesson materials, create quizzes, generate examples, and respond to student questions. Those capabilities may affect teaching workflows. They do not answer whether AI can replace the role of a tutor, whose value is created through diagnosis, responsive scaffolding, sustained attention, relationship, motivation, and accountability. A chatbot may produce a correct explanation. A tutor must decide why this student is stuck, whether the student is overconfident or anxious, which misconception is driving the error, whether the homework prompt conflicts with classroom expectations, when to intervene, and when to fade support.

For this reason, the paper uses a deliberately bounded definition: tutoring is individualized or small-group academic support that supplements, but does not replace, classroom instruction. The scope includes private tutoring, school-based high-impact tutoring, online tutoring, test-preparation support, subject-specific coaching, and hybrid models that combine human tutors with technology. The scope excludes classroom teaching, course design, school administration, and full-service instructional replacement. This framing is not semantic. It is the foundation for making sound policy and business decisions in an AI-driven education market.

The Evidence Base: Human Tutoring Produces Measurable Learning Gains

Human tutoring has long been among the most effective educational interventions studied. Bloom's well-known '2 sigma' analysis framed one-to-one tutoring as a benchmark for what individualized instruction can achieve, even as it challenged researchers to find scalable methods that approach those effects (Bloom, 1984). Later empirical syntheses found consistently positive results. Cohen, Kulik, and Kulik's meta-analysis of 65 evaluations reported that tutored students outperformed non-tutored control students on academic performance and attitudes toward the subject matter (Cohen et al., 1982). More recently, Nickow, Oreopoulos, and Quan reviewed randomized field experiments and found an average effect size of 0.288 standard deviations across 89 studies, with gains of 0.290 in literacy and 0.268 in mathematics (Nickow et al., 2024).

The implication is not that every tutoring program works equally well. The evidence favors tutoring designs that are frequent, structured, embedded in a student's academic life, and relationship-based. High-impact tutoring typically involves trained tutors, very small groups, multiple sessions per week, close monitoring of progress, and alignment with curriculum (National Student Support Accelerator, n.d.). These features distinguish tutoring from casual help, homework-answer services, or unstructured drop-in support. They also highlight why a purely technical substitute is difficult: the intervention is a service model, not just an explanation engine.

Tutoring demand also remains robust. In U.S. public schools, federal School Pulse Panel data showed that 87% of public schools offered some form of tutoring during the 2023-24 school year. The share offering high-dosage tutoring rose from 39% in October 2023 to 46% in May 2024, and schools using high-dosage tutoring rated it as at least moderately effective in large numbers (Institute of Education Sciences, 2024). Commercial data show parallel growth in online and private tutoring. Grand View Research estimated the global online tutoring services market at $10.42 billion in 2024 and projected it to reach $23.73 billion by 2030, a 14.5% compound annual growth rate. Its U.S. online private tutoring estimate was $4.33 billion in 2024, projected to reach $8.09 billion by 2030 (Grand View Research, 2025a, 2025b).

This evidence does not imply immunity from disruption. AI will change tutoring economics. It may reduce the price of basic explanation, automate content preparation, and intensify competition for low-end homework help. But the demand signal is not collapsing. If anything, demand is being reframed around measurable learning outcomes, program quality, affordability, and personalization. Human tutoring that cannot demonstrate learning value is vulnerable. Human tutoring that combines relationship, diagnostic skill, and effective technology is positioned to expand.

AI Adoption Is Real - But It Is Not the Same Demand as Tutoring

Generative AI has become a general academic utility. Among U.S. teens, Pew Research Center found that 57% used AI chatbots to search for information and 54% used them for schoolwork (McClain et al., 2026). College Board reported that high school students' use of generative AI for schoolwork rose from 79% in January 2025 to 84% in May 2025; common uses included brainstorming, editing and revising essays, conducting research, finding sources, and using ChatGPT for assignments or homework (College Board, 2025). At the university level, the Higher Education Policy Institute found that 92% of surveyed U.K. undergraduates used AI in some form in 2025, with many using it to explain concepts, summarize articles, and suggest research ideas (Freeman, 2025). A Harvard undergraduate survey similarly reported widespread use, including cases where students treated generative AI as a substitute for office hours or required readings (Hirabayashi et al., 2024).

These data are important because they describe a broad behavioral shift, not a simple substitution event. Most student AI use is immediate, self-serve, low-friction, and task-based. A student asks a chatbot to rephrase a paragraph, produce study questions, summarize a reading, or explain a homework step. That use may be beneficial, harmful, or academically questionable depending on context. But it is different from committing to a tutor over weeks or months, paying for recurring support, coordinating with parents or schools, and pursuing a defined learning plan.

The historic human tutoring market has been much narrower and more segmented than the new AI-use population. Kim, Goodman, and West documented that private tutoring centers in the United States tripled from roughly 3,000 to 10,000 between 1997 and 2022, with growth concentrated in high-income and high-parental-education communities. They also reported that only about 6% to 7% of U.S. families with children ages 6 to 17 paid for tutoring in 2022, even though spending among purchasers could be substantial (Kim et al., 2024). In other words, paid tutoring has historically been a selected market, not a universal academic behavior.

The contrast is stark. AI use for schoolwork is now reported by roughly half or more of teens and by very large shares of high school and college students, depending on population and measure. Paid tutoring has historically been used by a much smaller share of families and is shaped by income, academic need, enrichment goals, admissions competition, and parental preferences. These are not identical populations standing at a checkout screen choosing 'AI' instead of 'human tutor.' There is, of course, overlap: many tutored students also use AI, and some families may replace low-quality homework help with AI. But the dominant jobs-to-be-done differ. AI often solves an immediate task friction problem. Tutoring solves a learning, accountability, confidence, or advancement problem over time.

This distinction has strategic consequences. If an organization defines tutoring as 'answering student questions,' AI looks like a replacement. If tutoring is defined as a structured, relationship-based learning service, AI looks like infrastructure. It expands the volume of academic assistance students receive while creating new needs: students must learn how to question AI outputs, avoid overreliance, cite responsibly, preserve academic integrity, and convert AI-assisted performance into independent mastery. Those needs are squarely within the emerging value proposition of human tutoring.

What Human Tutors Do That AI Does Not Fully Replicate

The strongest tutoring effects do not arise simply because a learner receives more information. They arise from dynamic interaction. Classic research on scaffolding described tutoring as the process by which a more expert partner enables a learner to accomplish tasks that would otherwise be beyond independent reach, while gradually transferring responsibility back to the learner (Wood, Bruner, & Ross, 1976). Studies of human tutoring dialogues show that effective tutors adapt prompts, hints, questions, and explanations in response to what the student says and does (Chi et al., 2001; Graesser, Person, & Magliano, 1995). Expert tutors also manage motivation, persistence, and affect, not just content (Lepper, Drake, & O'Donnell-Johnson, 1997).

AI systems can imitate parts of this interaction. They can generate hints, identify common mistakes, produce alternate explanations, and maintain a conversational interface. But current AI does not fully replicate the situated responsibility of a human tutor. A human tutor observes hesitation, fatigue, frustration, avoidance, overconfidence, embarrassment, and effort. The tutor knows when a student is using an explanation as a shortcut, when a parent is asking for grades but the student needs confidence, and when a school assignment requires a specific method rather than a mathematically equivalent shortcut. The human tutor is accountable to the learner, the family, the institution, and professional judgment in a way that a general-purpose model is not.

Three forms of human value are especially difficult to automate. The first is diagnostic judgment. Students rarely present their true problem cleanly. 'I don't understand factoring' may mean weak multiplication fluency, anxiety after a poor quiz, a skipped prerequisite, confusion about notation, or a teacher's method that differs from the student's online examples. A tutor investigates the gap through conversation, error analysis, and observation. The second is relational accountability. The student often works harder because a person notices effort, remembers prior commitments, and cares whether progress happens. Research on student-teacher relationships links positive relationships with engagement and achievement, and high-impact tutoring models likewise emphasize sustained tutor-student relationships as a design feature (National Student Support Accelerator, n.d.; Roorda, Koomen, Spilt, & Oort, 2011). The third is metacognitive transfer. A tutor's goal is not to help the student complete today's problem with assistance; it is to help the student become more capable without assistance.

AI can weaken that transfer when used poorly. In a field experiment with high school mathematics students, Bastani and colleagues found that unrestricted GPT-4 access improved performance while students had access to the tool, but students performed worse than controls when AI access was later removed; a more guarded 'GPT Tutor' design mitigated the harm (Bastani et al., 2025). The lesson is not that AI is bad for learning. The lesson is that unguided AI can become a crutch. Human tutors are positioned to manage that risk by setting conditions for productive struggle, asking students to explain reasoning, checking independent performance, and fading support deliberately.

Finally, tutors provide ethical and institutional judgment. UNESCO has urged a human-centered approach to generative AI in education, emphasizing privacy, validation, pedagogical appropriateness, and the protection of human agency (Miao & Holmes, 2023). OECD guidance similarly highlights risks such as hallucinations, bias, privacy concerns, and overreliance (OECD.AI Policy Observatory, 2024). These risks show why academic AI use needs human interpretation. A tutor can help a student ask better questions, verify sources, distinguish explanation from answer-generation, comply with assignment rules, and understand why integrity matters. A model can output content; a tutor can coach judgment.

AI as an Augmentation Layer for Tutoring

The case for human tutoring does not require minimizing AI. On the contrary, the most compelling future is one in which tutors become more effective because they use AI well. Intelligent tutoring systems and AI-supported help have already shown learning potential. VanLehn's review found that human tutoring and intelligent tutoring systems had similar average effects in the studies reviewed, with human tutoring at d = 0.79 and intelligent tutoring systems at d = 0.76 (VanLehn, 2011). Pardos and Bhandari compared ChatGPT-generated help with human-tutor-authored help across mathematics skills and found that AI-generated help could produce learning gains comparable to human-authored help, though the model's initial outputs failed quality checks on 32% of problems before mitigation (Pardos & Bhandari, 2024). Kestin and colleagues reported that a purpose-built AI tutor improved learning in an undergraduate physics setting relative to an active-learning comparison condition (Kestin et al., 2025).

These studies point toward augmentation rather than replacement because they show both capability and constraint. AI can generate useful instructional support when the domain is bounded, the prompts are designed carefully, and quality checks exist. It can reduce preparation time, provide many examples, translate explanations into different reading levels, and generate practice opportunities immediately. But the same evidence base also shows the need for guardrails, validation, and alignment to learning goals. A human tutor can turn AI from an answer generator into a learning instrument.

School-based experimentation is already moving in this direction. The University of Chicago Education Lab reported that a Saga Education model combining tutor time with education technology reduced costs by about one-third and cut the number of tutors required roughly in half without a drop-off in effectiveness; students in the model achieved the equivalent of an additional one to two years of math learning (University of Chicago Education Lab, 2024). The strategic significance is clear: technology can make tutoring more scalable when it is embedded in a coherent human service model. It does not have to remove the tutor from the loop.

In practice, AI can augment tutors across the tutoring workflow. Before a session, a tutor can use AI to transform assessment data into hypotheses about misconceptions, generate warm-up items aligned to a student's current unit, or create multiple versions of a practice set. During a session, AI can supply alternative explanations, worked examples, simulations, vocabulary support, or bilingual scaffolds while the tutor observes reasoning and chooses which path to use. After a session, AI can help summarize progress, draft parent updates, tag skill gaps, and recommend independent practice. Across a program, AI can support quality assurance by flagging patterns in student work and identifying when a student may need a different intervention.

The tutor's role changes under this model. The tutor is no longer the scarce source of every explanation or worksheet. The tutor becomes the orchestrator of a learning environment: selecting the right task, managing cognitive load, asking the next question, ensuring the student articulates reasoning, and protecting independence. This is a higher-skill role, not a lower-skill one. The more AI can produce plausible answers, the more valuable it becomes for a human to determine whether a student has learned.

AI Creates New Gaps That Tutors Are Well Positioned to Fill

AI does not only automate existing academic support. It creates new learning gaps. Students now need AI literacy: how to write useful prompts, evaluate model outputs, detect hallucinations, cross-check evidence, cite appropriately, and understand when AI assistance violates an assignment's purpose. These skills are not peripheral. They are becoming part of academic competence. Yet student adoption has outpaced institutional support. HEPI found that only 36% of surveyed students said they had received support from their institution for developing AI skills, despite very high AI use for assessments (Freeman, 2025).

Tutors can fill this gap because they meet students at the moment of use. A classroom policy may state what is allowed, but a tutor can help a student decide what to do with a real essay prompt or problem set. A tutor can say: use AI to generate three possible study questions, but solve the problem yourself; ask AI to critique your thesis, but verify its claims; request an explanation at a lower reading level, then close the tool and teach the concept back; compare the model's solution method with your teacher's method; identify which part you understand and which part you are borrowing. This coaching turns AI from a productivity shortcut into a metacognitive mirror.

The biggest opportunity is not to ban AI from tutoring, but to design tutoring around responsible AI use. Tutors can help students develop habits of productive struggle: attempting first, using AI for hints rather than answers, explaining reasoning aloud, checking work independently, and reflecting on what changed. Tutors can also protect equity. Students with highly educated parents may receive sophisticated guidance on AI use at home; students without that support may use AI in riskier, less educational ways. AI-literate tutors can democratize access to better academic decision-making.

This opportunity also strengthens the business case for tutoring providers. Providers can offer AI-integrated learning plans, responsible-use coaching, skill diagnostics, progress reporting, and tutor training as differentiated services. Rather than competing with free chatbots on answer production, they can compete on outcomes, trust, safety, and student independence. In an AI-saturated market, those are the scarce assets.

Strategic Implications for Educators, Policymakers, and EdTech Stakeholders

For schools and districts, the implication is to treat AI as tutoring infrastructure, not as a wholesale substitute for tutors. High-impact tutoring remains a promising use of resources when designed around frequency, small groups, curriculum alignment, trained adults, and progress monitoring. AI can reduce administrative and content-preparation burden, but it should not remove the human relationships that make tutoring accountable. Districts should evaluate AI-enabled tutoring models by independent learning gains, not by engagement minutes or answer accuracy alone.

For tutoring providers, the imperative is professionalization. Tutors need training in AI literacy, academic integrity, data privacy, prompt design, and evidence-based scaffolding. Providers should create clear protocols for when tutors may use AI, how outputs are reviewed, how student data are protected, and how sessions verify independent understanding. The provider that merely offers more practice questions will face price pressure. The provider that combines human judgment, responsible AI, and measurable progress can justify premium value.

For edtech developers, the opportunity is to build tutor co-pilots rather than student answer machines. Tools should help tutors diagnose misconceptions, generate aligned practice, review student work, produce formative insights, and document progress. They should make it easier for a human to see what the student understands, not easier for a student to bypass understanding. Product metrics should therefore include independent post-assessment performance, reduction in tutor prep time, quality of tutor decisions, and evidence of transfer.

For policymakers and funders, the core issue is equity. Historically, private tutoring access has been uneven and income-skewed (Kim et al., 2024). AI could widen or narrow that gap. It may provide free academic help to many students who never had tutors. But without human support, it may also deepen differences in how well students use the technology. Public investment should favor human-in-the-loop models that expand access to high-impact tutoring, require evidence of learning, protect privacy, and make responsible AI guidance available to students who would otherwise lack it.

Design Principles for the AI-Augmented Tutor

The human tutor remains essential when AI is designed around learning rather than answer delivery. Five principles should guide implementation:

Human accountability: A named tutor or program should be responsible for interpreting AI outputs, monitoring progress, and deciding when a student has achieved independent mastery.
Human review: AI-generated explanations, practice items, summaries, and assessments should be reviewed by trained adults before they shape high-stakes decisions or reinforce misconceptions.
Privacy by design: Student data should be minimized, protected, and governed by clear consent and retention rules. AI tools should not become informal repositories of sensitive educational records.
Evidence of learning: Programs should measure transfer to independent work, not just completion, satisfaction, or time on task. The key question is whether the student performs better when the scaffolds are removed.
Equity: AI-enabled tutoring should lower access barriers without lowering quality. The aim is not a cheaper version of weak tutoring; it is wider access to the features that make tutoring effective.

The Scarce Resource Is Human Judgment

AI will transform academic assistance. It will make explanations cheaper, practice more abundant, and learning tools more available. It will also create new risks: plausible misinformation, overreliance, hidden inequities, and confusion about what counts as learning. Those changes do not eliminate the need for human tutoring. They clarify it.

The tutor's enduring value is not that humans can recite facts better than machines. They cannot. The tutor's value is that learning is a human process involving confidence, identity, judgment, motivation, context, and transfer. A tutor notices what a student avoids, remembers what the student promised to practice, asks the inconvenient follow-up question, interprets the teacher's expectations, helps a family understand progress, and insists that assisted performance become independent capability.

The future, then, is not human tutoring versus AI. It is weak tutoring versus AI-augmented tutoring. Weak tutoring that merely supplies answers or generic explanations will be displaced. Strong tutoring will become more data-informed, more personalized, more affordable, and more central to helping students navigate an AI-rich academic world. Human tutors will remain indispensable not because AI is unimportant, but because AI makes human judgment, relationship, and accountability more important.

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O., & 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. https://doi.org/10.1073/pnas.2422633122

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. https://doi.org/10.3102/0013189X013006004

Chi, M. T. H., Siler, S. A., Jeong, H., Yamauchi, T., & Hausmann, R. G. (2001). Learning from human tutoring. Cognitive Science, 25(4), 471-533. https://doi.org/10.1207/s15516709cog2504_1

Cohen, P. A., Kulik, J. A., & Kulik, C.-L. C. (1982). Educational outcomes of tutoring: A meta-analysis of findings. American Educational Research Journal, 19(2), 237-248. https://doi.org/10.3102/00028312019002237

College Board. (2025, October 6). New research: Majority of high school students use generative AI for schoolwork. College Board Newsroom.

Freeman, J. (2025). Student Generative AI Survey 2025. HEPI Policy Note 61. Higher Education Policy Institute.

Graesser, A. C., Person, N. K., & Magliano, J. P. (1995). Collaborative dialogue patterns in naturalistic one-to-one tutoring. Applied Cognitive Psychology, 9(6), 495-522. https://doi.org/10.1002/acp.2350090604

Grand View Research. (2025a). Online tutoring services market size, share & trends analysis report, 2030.

Grand View Research. (2025b). U.S. online private tutoring market size, share & trends analysis report, 2030.

Hirabayashi, S., Jain, R., Jurkovic, N., & Wu, G. (2024). Harvard undergraduate survey on generative AI. arXiv. https://arxiv.org/abs/2406.00833

Institute of Education Sciences. (2024, July 18). About one-quarter of public schools reported that lack of focus or inattention from students had a severe negative impact on learning in 2023-24. National Center for Education Statistics.

Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT in an undergraduate physics course. Scientific Reports, 15, 17458.

Kim, E., Goodman, J., & West, M. R. (2024). Kumon In: The recent, rapid rise of private tutoring centers. EdWorkingPaper No. 21-367. Annenberg Institute at Brown University. https://doi.org/10.26300/z79x-mr65

Lepper, M. R., Drake, M. F., & O'Donnell-Johnson, T. (1997). Scaffolding techniques of expert human tutors. In K. Hogan & M. Pressley (Eds.), Scaffolding student learning: Instructional approaches and issues. Brookline Books.

McClain, C., Anderson, M., Sidoti, O., & Bishop, W. (2026). How teens use and view AI. Pew Research Center.

Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO.

National Student Support Accelerator. (n.d.). What is high-impact tutoring? Stanford University.

Nickow, A. J., Oreopoulos, P., & Quan, V. (2024). The promise of tutoring for PreK-12 learning: A systematic review and meta-analysis of the experimental evidence. American Educational Research Journal, 61(1), 74-107. https://doi.org/10.3102/00028312231208687

OECD.AI Policy Observatory. (2024). Generative AI: The risks and unknowns. Organisation for Economic Co-operation and Development.

Pardos, Z. A., & Bhandari, S. (2024). ChatGPT-generated help produces learning gains equivalent to human tutor-authored help on mathematics skills. PLOS ONE, 19(5), e0304013. https://doi.org/10.1371/journal.pone.0304013

Roorda, D. L., Koomen, H. M. Y., Spilt, J. L., & Oort, F. J. (2011). The influence of affective teacher-student relationships on students' school engagement and achievement: A meta-analytic approach. Review of Educational Research, 81(4), 493-529. https://doi.org/10.3102/0034654311421793

University of Chicago Education Lab. (2024, June 3). Study finds in-school high-dosage tutoring combining technology and tutor time has no drop-off in effectiveness, reduces costs.

VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197-221. https://doi.org/10.1080/00461520.2011.611369

Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89-100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x