AI tutoring is most effective when it assists human teachers rather than working alone with students. A new research brief from Stanford University's AI Hub for Education examined emerging evidence on AI tutoring models and found that the strongest results come from tools designed to help teachers do their jobs better, not from chatbots that interact directly with students.

What Does the Research Actually Say About AI Tutors?

The challenge with understanding AI tutoring effectiveness is that the term covers a wide range of very different approaches. "People hear the term 'AI tutor,' and they think that is a young person engaging directly with a chatbot," explained Chris Agnew, director of the AI Hub for Education at Stanford's SCALE Initiative. "What our brief outlines is that AI tutors exist on an AI-human spectrum, and that spectrum exists on a range of relational intensity. Depending on where you fall on that range, there is either strong evidence to support it or very little evidence".

"Tutoring research tells us that relationships are a key part in student persistence, student engagement, and hitting dosage that allow sustained improvement in student outcomes. Access to a tool is not enough," said Chris Agnew.

Chris Agnew, Director of the AI Hub for Education at Stanford's SCALE Initiative

The Stanford research brief categorized tutoring models by how much human involvement they include, then mapped them against what research shows actually works:

• In-person or remote tutoring: A human tutor handles all instruction and student interaction. This model has robust evidence showing it can be highly effective.
• Human tutoring with AI support: A human tutor interacts with the student while an AI tool assists behind the scenes. Research is still emerging but suggests this could be as effective or more effective than human-only tutoring.
• AI tutoring with human support: A student engages directly with AI while a human tutor oversees, guides, and intervenes when needed. Research is emerging and indicates that implementation determines whether it works.
• AI-only tutoring: A student works directly with AI without direct human oversight. There is limited research on this model; one study found that 40 to 47 percent of students never used the AI platform when left to work independently, making it difficult to measure effectiveness.

The findings challenge the assumption that AI can simply replace human tutors. When students work alone with AI tutoring systems, engagement drops dramatically. "Many students don't engage with AI tutors," the research brief found. Adding human oversight helped with engagement, but the strongest evidence for actual learning gains came from AI tools built for human tutors to use.

How Can Schools Use AI Tutoring Effectively?

Rather than deploying AI chatbots as standalone tutors, schools can use AI to enhance their existing tutoring programs and free up teacher time for what humans do best. The Stanford researchers recommend several practical applications:

• Master scheduling: Use AI to streamline student and educator allocation, which can reduce schedule-building time and preserve core instructional time and intervention blocks.
• Tutor training: Enhance tutor professional development through realistic practice simulations that allow tutors to hone their instructional techniques before leading live sessions.
• Practice materials: Generate targeted student practice materials with a tutor reviewing them before implementation.

Matthew Kraft, a professor of education and economics at Brown University, noted that while personalized learning can succeed without direct human relationships, "evidence from more structured computer-adaptive learning software programs is promising when schools can structure implementation successfully". However, he cautioned that "AI cannot ensure you meet the basic elements of successful tutoring: high attendance, sustained tutoring over time, and strong relationships".

Why Is England Treating Schools as AI Testing Laboratories?

Meanwhile, England is taking a different approach to understanding AI tutoring at scale. The Department for Education has launched an "AI Tutoring Tools Pioneers Programme" and an edtech "testbed" scheme that will recruit schools nationwide as AI testing sites beginning in September 2026. This represents a deliberate strategy to generate evidence about AI tutoring by testing prototype applications in real classrooms.

The English government's approach reflects what researchers call "experimental futuring," a method that uses rapid product development and evaluation trials to shape policy. The Department for Education worked with the company FacultyAI to create proof-of-concept AI prototypes for teaching, then seed-funded edtech companies to further develop and test applications based on standardized curriculum materials and teaching standards.

Education Secretary Bridget Phillipson announced at the BETT edtech trade show in January 2026 that the government would invest an additional 23 million pounds to expand the EdTech Testbed pilot into a four-year program. "It'll recruit schools and colleges to put the latest tech and AI tools through their paces, in the cut and thrust of classrooms across the country," she stated.

This testbed approach treats schools as what researchers call "living laboratories" for evidence production on previously untested prototypes. Rather than waiting for perfect evidence before deploying AI tools, the government is generating evidence through live testing in real school settings. The goal is to identify which AI tutoring approaches show the most promise before scaling them more widely.

The key takeaway from both the Stanford research and England's experimental approach is the same: AI tutoring is not a simple replacement for human instruction. Its value lies in how it augments human teaching, reduces administrative burden, and creates opportunities for more personalized learning when implemented thoughtfully. Schools considering AI tutoring should prioritize tools that enhance human-led instruction and establish strict data-privacy safeguards before adoption.