The AI Tutor Is Already Substituting, Not Supplementing

When Khan Academy launched Khanmigo, its AI tutoring system, in 2023, the framing was supplement, not substitute. The system would help students who needed additional practice, provide personalized explanations outside classroom hours, and free teachers to focus on the higher-order aspects of their work. This framing was reassuring to teachers, to policy makers, and to the public. It was also, within approximately 18 months, demonstrably incorrect.

The data from Khanmigo and the broader ecosystem of AI tutoring platforms that followed it tells a different story — one that the developers of these platforms have been careful not to publicize, that the schools adopting them have not fully processed, and that the teacher unions have begun to notice but have not yet translated into coherent political response.

The Signal

RAND Corporation's 2025 analysis of AI tutoring platform usage data across 340 school districts found that in districts where AI tutoring systems were deployed, student-initiated interactions with AI tutors exceeded student-initiated interactions with human teachers in mathematics by a ratio of 3.1 to 1 after 18 months of deployment. In reading comprehension, the ratio was 1.7 to 1. In science, it was 1.2 to 1.

This is not the supplementation pattern the platforms promised. When students are going to the AI tutor more than three times as often as they go to their teacher for help with math, the AI tutor is not a supplement — it is the primary instructional relationship for a significant fraction of learning interactions. The teacher has become the secondary resource.

The quality of these interactions is a separate question from the quantity — and the quality data is more mixed. But the substitution pattern is real and structural: it reflects student preferences, efficiency calculations, and the availability characteristics of AI systems (always available, never impatient, infinitely patient with repeated questions) that human teachers cannot match and should not try to.

The Historical Context

The history of educational technology is a history of tools that were adopted as supplements and remained supplements. The textbook, the filmstrip, the overhead projector, the computer lab, the internet — each was introduced with claims that it would transform teaching, and each was absorbed into the existing institutional structure of schools without fundamentally restructuring the teacher-student relationship. The educational research community developed a somewhat weary consensus: technology in education tends to complement existing practices rather than disrupt them.

The AI tutor is different in a structural way that the prior examples were not. The textbook, the filmstrip, and the computer lab could not engage in dynamic, personalized dialogue. They were resources to be consulted, not relationships to be maintained. The AI tutor is, for practical purposes, a dynamic conversational partner that can respond to the specific misconception of the specific student in real time, provide as many alternative explanations as needed, and never lose patience or attention.

The prior technology that is most analogous is not a classroom tool but a communications technology: the telephone, and subsequently the internet, which substituted for face-to-face interaction in ways that prior communications technologies could not. These technologies did not supplement human communication — they substituted for significant fractions of it. The AI tutor is doing the same thing to educational interaction.

The Mechanism

Three characteristics of AI tutoring systems explain the substitution pattern.

Availability asymmetry is the most basic. A teacher with 28 students is available for approximately 90 seconds per student per class period for individualized interaction. An AI tutor is available for unlimited individualized interaction at any time. For students who have questions outside classroom hours, or who need more time with a concept than the class schedule allows, the substitution is rational: the AI tutor is simply the available option.

Patience and non-judgment characteristics reduce the social cost of asking repetitive questions. Many students who would not ask a teacher to explain something a third or fourth time will ask an AI tutor repeatedly. The social friction of human interaction — the embarrassment of appearing slow, the imposition on the teacher's limited time — is absent in AI interaction. This makes the AI tutor more accessible for the interactions that students most need and are least likely to seek from human teachers.

Personalization at scale addresses the fundamental constraint of classroom teaching. A teacher covering thirty students must calibrate instruction to the median student's needs, which means that the students who are above and below the median are systematically underserved. AI tutors can operate at the exact level of the individual student and adjust in real time. This is not a marginal improvement — it addresses the structural limitation that has constrained educational quality since schooling became compulsory.

Second-Order Effects

The teaching profession faces a structural disruption that is different in kind from prior technology-driven changes to educational practice. Prior technologies changed what teachers did with the time they spent with students; AI tutoring changes how much instructional time students seek from teachers at all. If the RAND substitution ratios continue to compound, the number of teacher-student instructional interactions per student per year will decline significantly — not because teachers are being replaced, but because students are choosing the AI tutor for an increasing share of interactions.

This has implications for teacher workforce sizing that are not yet in the policy conversation. If AI tutors handle 50-60% of individualized instructional interactions, the student-to-teacher ratio that produces equivalent learning outcomes can increase. This is the mechanism through which AI tutoring potentially reduces the teacher workforce — not by replacing individual teachers but by allowing each teacher to effectively serve more students.

The learning outcome data for AI-tutored students versus teacher-tutored students in equivalent subjects is not yet sufficiently mature to assess whether the quality of instruction is comparable. The preliminary studies suggest that AI tutoring produces equivalent or better outcomes in procedural subjects (mathematics, grammar) and worse outcomes in subjects requiring interpretive judgment (literary analysis, historical argument). If this pattern holds in larger studies, the substitution may be rational where it is occurring (STEM subjects) while problematic in the humanities.

What to Watch

Longitudinal learning outcome studies: Randomized controlled trials comparing AI-tutored and non-AI-tutored students over multiple years will provide the definitive evidence on learning outcomes. Watch for results from the major educational research institutions.

Teacher union response: The substitution pattern creates a genuine workforce threat that unions will eventually respond to. Watch for collective bargaining language addressing AI tutoring deployment and student-to-teacher ratio adjustments.

District AI tutoring adoption rates: The fraction of US school districts deploying AI tutoring systems is currently 23% and growing. If it crosses 50%, the substitution pattern becomes a national educational structure issue rather than a local policy question.

Subject-specific outcome divergence: If AI tutoring platforms begin publishing subject-specific learning outcome data, watch for the pattern of STEM advantage and humanities disadvantage that preliminary studies suggest. This would clarify where substitution is beneficial and where it is harmful.