AI can explain a math problem in seconds.
It can generate practice questions.
It can walk through an equation.
It can summarize a concept ten different ways until one finally clicks.
And it is available at 10:30 at night when no tutor is.
So the temptation is easy to understand:
Why pay for a human tutor at all?
A new two-year study of AI tutoring gives families a useful answer.
AI can help students learn. But access to an AI tutor is not the same thing as having an effective tutor. And for students trying to move from good to exceptional performance, that distinction matters more than most families realize.
What the Research Actually Found
The August 2026 NBER working paper One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment followed students across 18 Tennessee middle schools over two years. Students were given Khan Academy with Khanmigo — its generative-AI tutor — during existing remedial math periods. Khanmigo was deliberately configured to coach students rather than simply give them answers.
The results are instructive.
Students improved.
But the researchers found that the improvement looked surprisingly similar to what would have been expected from structured Khan Academy practice without the AI tutor.
The AI was there. Students used it. And it largely did not move the needle beyond what regular practice would have done anyway.
That is the finding families should understand.
The AI Was Available. Students Barely Used It.
Almost every student tried Khanmigo at least once — 96% used it.
Sounds promising.
But then look at what happened next.
The median student interacted with Khanmigo on only about one-third of the days they practiced. Students used it in only 17% of the exercise sessions where they had made a mistake. And only 14.5% of student messages contained an actual mathematical question or a step of reasoning.
Think about that.
The student gets something wrong. A personal AI tutor is sitting right there. And more than 80% of the time, the student does not meaningfully engage it.
That is the problem families should understand.
Having help available is not the same as knowing when you need help, wanting help, or knowing what question to ask.
And those are exactly the moments when a great human tutor becomes valuable.
Students Don’t Know What They Don’t Know
This is one of the fundamental problems with asking a student to manage their own tutoring.
The student has to recognize: “I don’t understand this.”
Then they have to identify what they do not understand.
Then they have to formulate the right question.
Then they have to remain engaged long enough to work through the answer.
That sounds simple. It often is not.
A student may think: I just made a careless mistake.
A skilled tutor may see: No. You have made this same mistake four times because you fundamentally misunderstand what the question is asking.
Those are very different diagnoses.
AI can be remarkably good at responding to a question. The harder problem is that the student may never ask the question that actually needs to be asked.
The researchers saw exactly this. They concluded that the limiting factor was not what the AI was capable of doing, but whether students actually engaged with it in meaningful ways. Even with free access, mandatory practice time, teacher supervision, and an interface encouraging students to use the tutor, substantive engagement remained low.
A Great Tutor Does Not Wait for the Student to Ask for Help
This is where human tutoring works differently.
A good tutor is watching. Listening. Questioning. Looking for hesitation. Looking for patterns. Looking at how the student arrives at the answer, not simply whether the answer is correct.
A student can answer a question correctly and still reveal a problem.
Maybe the approach took three minutes when it should have taken thirty seconds. Maybe the student guessed between two choices. Maybe the shortcut happened to work this time but will fail when the question changes. Maybe the student memorized a process without understanding why it works.
Software often sees: Correct. Next question.
A skilled tutor asks: “Why did you do it that way?”
That question can reveal more than ten additional practice problems.
The paper itself highlights why individualized human tutoring has historically been so powerful: a tutor can diagnose a specific error, explain the missing step, and adapt what comes next to the student’s particular misunderstanding. Research cited by the authors puts the effects of strong one-on-one tutoring among the largest educational interventions ever studied.
AI Has Another Problem: It Can Make Learning Feel Easier Than It Is
There is a subtler trap.
AI is extraordinarily good at explaining things.
A student asks a question. AI explains it. The explanation makes perfect sense. And the student thinks: Got it.
But did they actually learn it? Or did they simply understand the explanation while it was sitting in front of them?
Those are not the same thing.
This connects directly to why passive instruction fails high-achieving students. Watching someone solve a problem — or reading a clear AI explanation — creates recognition. Real mastery requires the student to generate the reasoning independently.
The NBER paper discusses an earlier experiment in which high school students given unrestricted GPT-4 performed better while they had AI available — but performed worse on a later exam when the AI was removed. A more carefully restricted tutoring version that guided students instead of simply doing the work avoided that negative effect.
That is an important distinction.
Getting more questions right while practicing is not necessarily the same thing as learning more.
For SAT and ACT preparation, that matters enormously. Because AI does not come into the testing room.
The Real Test Is What Happens When the Help Disappears
Imagine a student practicing SAT math.
They encounter a difficult problem. They paste it into an AI tool. The AI identifies the concept, explains the setup, points out the shortcut, and walks them toward the answer.
Excellent.
Now give the student a different question testing the same underlying concept. Change the wording. Hide the variable. Use a different diagram. Remove the obvious clue.
Can the student still solve it?
That is the question that matters.
The SAT and ACT reward transfer — the ability to recognize a familiar concept inside an unfamiliar question. Students must independently decide: What is this really testing? Which information matters? What strategy should I use? Am I falling into a trap?
If the student has become dependent on AI to identify the pattern first, they may be practicing the wrong skill. They are practicing asking for help rather than recognizing the problem themselves.
AI Gives Students What They Ask For. A Tutor Gives Them What They Need.
Here is one of the clearest differences between AI and human tutoring.
A student says: “Can you show me how to do this?”
A good tutor might say: “No. Show me what you tried first.”
The student says: “I don’t know.”
The tutor says: “Start anyway.”
That may feel less efficient. Educationally, it is often far more valuable.
Because productive struggle matters. The goal is not to make every practice question easy. The goal is to make the student more capable when the questions are hard.
AI is very good at removing friction. That is sometimes exactly what learning does not need.
This Matters Even More for Top Scorers
The students Crownridge Coaching works with are rarely trying to move from the 50th percentile to the 60th percentile.
They may be trying to move from:
- a 1400 to a 1500+
- a 1450 to a 1550+
- a 32 to a 34 or higher
- a strong PSAT score into National Merit territory
At that level, broad explanations are usually not the problem. The student often already knows the content. The remaining points disappear because of narrow, specific issues:
- one recurring reasoning mistake
- inefficient problem selection under time pressure
- subtle reading traps
- overthinking a question type they actually know
- failing to recognize a familiar concept presented differently
- abandoning a correct approach before completing it
Those problems are difficult for a student to self-diagnose. And an AI tutor depends entirely on the student initiating the interaction and communicating enough information for the system to identify what is happening.
The NBER researchers explicitly noted that even when students made mistakes, meaningful AI dialogue remained uncommon. They concluded that human attention may be an important ingredient in realizing the full potential of personalized learning.
That is particularly relevant to advanced test prep. The last 50 or 100 SAT points may not require another lesson on algebra. They may require someone to notice something about the student that the student cannot notice about themselves.
What the Study Does and Doesn’t Show
This research should not be overstated.
The study involved middle-school students in remedial mathematics — not high-achieving juniors preparing for the SAT or ACT. The authors acknowledge that AI technology is evolving and that future systems with better awareness of a student’s work and more proactive engagement may produce different results. The paper is also an NBER working paper, circulated for discussion rather than peer-reviewed.
So the conclusion is not: “AI cannot help students learn.”
The conclusion is: access to AI tutoring and the active benefits of AI tutoring are two very different things — and the gap between them is larger than most families expect.
That finding is unlikely to change dramatically just because the underlying technology improves. The engagement problem, the self-diagnosis problem, and the productive-struggle problem are not primarily technology limitations. They are human ones.
Where AI Has a Legitimate Role
None of this means students should avoid AI tools.
Used in the right role, they can be genuinely useful for reinforcement — generating additional practice questions, explaining a step from a different angle, quizzing on material already introduced, or helping a student review a concept between sessions.
That is a meaningful supporting function.
What it is not is a substitute for the diagnostic work, the proactive observation, and the calibrated pressure that a skilled human tutor provides.
There is a significant difference between:
“I work with a tutor, and I use AI tools between sessions to reinforce what we covered.”
and:
“AI is my tutor.”
The first is a sensible combination. The second puts the entire burden of recognizing weaknesses, diagnosing misconceptions, resisting shortcuts, and creating accountability on the student — which is precisely the work they most need help with.
What Parents Should Watch For
If your student is using AI for school or test preparation, do not just ask: “Are you using it?”
Ask: “How are you using it?”
There is a significant difference between:
“Solve this SAT problem.”
and:
“Don’t give me the answer. Ask me questions that help me figure out where my reasoning went wrong.”
Between:
“Explain this concept.”
and:
“Make me explain this concept back to you, then challenge anything I don’t fully understand.”
One creates dependence. The other can build genuine capability.
The distinction matters even more for PSAT, SAT, and ACT preparation — where the student will ultimately sit alone in a room with no AI, no hints, and a clock running.
Final Thought
The NBER study is a useful reality check for families evaluating their options.
Students with access to an AI tutor did improve. But the gains looked similar to those from structured practice without AI, and most students rarely engaged the tutor deeply enough for conversational tutoring to explain much of the improvement.
The technology was there. The missing ingredient was the interaction.
That is something skilled tutors have understood for a long time. Learning does not happen because someone — or something — knows the answer. Learning happens because the student is made to think.
AI tools can play a supporting role in that process.
But for students chasing exceptional SAT and ACT results, the tutor running the process should be a person who can watch, diagnose, challenge, and adapt — not a system waiting to be asked the right question.
