The Real Question About AI Tutoring

A new analysis from Georgetown University's FutureEd lays out two competing visions for artificial intelligence in tutoring, and the gap between them says more about pedagogy than about technology. On one side sits a model that uses AI to stand in for the tutor. On the other sits a model that uses AI to make a human tutor sharper. The industry has turned this into a debate about cost versus relationships. That is the wrong argument.

Two Roads, One False Choice

The first approach, illustrated by Google's LearnLM, is a substitution model. The AI generates a response to a student, and a supervising tutor approves or edits it before it goes out. In a study of 165 students ages 13 to 15, tutors accepted 76 percent of the model's responses with little or no editing. Students working with LearnLM reached a 66 percent success rate on challenging follow-up topics, compared with 61 percent for students tutored by humans alone. The appeal is obvious: fewer tutor hours per student, and lower cost at scale.

The second approach, Tutor CoPilot, is a supplementary model. Here the AI offers a human tutor three suggested responses, and the tutor chooses, edits, or regenerates. In a study of roughly 1,000 elementary students, learners were four percentage points more likely to reach topic mastery when their tutor used the tool. The cost was about 20 dollars per tutor, mostly in reduced training time. The human stays in the room. The AI sits behind the human's shoulder.

Framed this way, schools appear to face a stark trade-off: buy scale and lose the relationship, or protect the relationship and pay for it. But hidden inside the second study is a finding that reframes the entire conversation.

The Number That Actually Matters

Tutors using CoPilot were 10 percentage points more likely to prompt students to explain their own thinking. That is the mechanism. Not the suggestions themselves, not the cost savings, but the shift in what the tutor asked the student to do. The tool nudged adults away from delivering answers and toward making students reason out loud.

Once you see that, the substitution-versus-support debate loses its edge. A supplementary tool can still let a student go passive if the tutor simply reads the AI suggestion aloud. And a substitution tool can keep a student fully engaged if it is built to ask rather than to tell. The delivery method is not destiny. The design is.

Substitution Is Not the Enemy. Passivity Is.

The fear underneath the whole conversation is that automated tutoring produces students who can prompt a machine but cannot reason on their own. That fear is legitimate. But it is not a fear of automation. It is a fear of answer-vending: any system, human or machine, that resolves the difficulty before the learner has wrestled with it.

The Socratic tradition solved this problem 2,400 years ago, and the solution transfers cleanly to software. A well-built AI tutor does not hand over the answer. It keeps the student in question space long enough to build the reasoning that transfers to the next problem. In practice, that means a system designed to do the following:

"Generative AI can reshape tutoring by serving more students at lower cost, though tradeoffs exist between scalability and the relational elements crucial to learning." — Tara Moon, FutureEd

What This Means for Schools Choosing Tools

For a community college worried about developmental-education completion, or a district stretching a handful of tutors across hundreds of students, cost and scale are real constraints, not luxuries. Substitution models will be part of the answer, because the alternative for many students is no tutoring at all. The task is not to reject automation on principle. The task is to demand that automated tutoring preserve the one thing that makes tutoring work.

This is the standard we built the Core-LX AI platform against. Socrat, our AI tutor, uses Socratic questioning to keep students reasoning rather than receiving, adjusts to each student's reading level while holding the rigor constant, and tracks mastery continuously so it can tell effort from understanding. It deploys inside an existing LMS through LTI in hours, and it supports more than 150 languages, so the students most often left out of high-quality tutoring are the ones it is designed to reach first. We are not claiming outcomes we have not measured; we are running a research pilot precisely to measure them.

So when a vendor pitches an AI tutor, the question to ask is not whether a human is in the loop or how cheaply it scales. Ask this instead: after a session, is the student able to do the next problem without the tool? If the honest answer is no, it is not a tutor. It is a very expensive way to copy answers.

Read the Full Article

Read "Two Emerging Strategies for Using AI in Tutoring" on FutureEd

About Core Learning Exchange: We provide turnkey Career and Technical Education (CTE) solutions for grades 6-14, offering 450+ courses from 20+ providers aligned to state standards and industry certifications. Our AI platform uses proven Socratic methodology to develop critical thinking skills through personalized, adaptive learning—deployed in hours via LTI integration.

Related Posts