The Quiet Collapse of the AI Tutor Dream
What Khanmigo reveals about why AI alone won’t fix learning
For the past two years, we have been told that AI tutors would change everything.
Students would have constant support. Learning would be personalized. Gaps would close.
And for a moment, it felt possible.
Then something quieter happened.
Students stopped using them.
What Happened to Khanmigo
When Sal Khan introduced Khanmigo through Khan Academy, the vision was clear. This would not be a chatbot that gives answers. It would guide students through thinking.
A tutor that prompts instead of solves.
A system built on questions, not shortcuts.
But more recently, that vision has run into a classroom reality.
Students often did not engage with it in meaningful ways. In many schools, it became a tool that was available but not deeply used.
That is not a technical failure. It is something deeper.
A Classroom Moment We Should Recognize
A student opens an AI tutor, pastes the question, and waits.
The tool responds with guiding prompts.
The student skips them.
They push for a more direct answer, copy what they need, and move on.
The assignment gets completed.
The learning does not.
That moment is not rare. It is the pattern that is happening all the time in our classrooms.
The Problem No One Wanted to Name
AI tutors depend on behaviors many students are still developing.
Persistence
Curiosity
Willingness to struggle
Ability to ask better questions
Khanmigo’s design assumes students will engage in that kind of thinking.
Many do not.
And without that engagement, the system has nothing to work with.
What the Research Actually Says
The evidence is more consistent than the headlines.
Studies show AI tutors can:
Improve performance during guided practice
Provide immediate feedback at scale
Support revision and iteration
This aligns with established approaches like formative assessment and mastery learning.
But the same research raises concerns.
Students using AI tools often show:
Limited reflection
Weak transfer of knowledge
Lower performance when working independently
In some studies, students completed tasks more successfully with AI, but performed worse on assessments without it.
That gap matters.
Some Research To Look At:
Brooking - What the research shows about generative AI in tutoring
Journal of Teaching and Learning - Leveraging “Khanmigo” Generative AI-Powered Tool for Personalized Tutoring to Learn Scientific Concepts.
What Teachers Are Actually Seeing
In classrooms, AI tutors are not replacing instruction. They are being repositioned.
Teachers are using them:
As practice tools, not primary teaching tools
Within structured assignments, not open-ended exploration
With clear guardrails, not full autonomy
And something else is happening.
Assignments are changing.
If AI can complete the task easily, the task is being reconsidered.
If AI can complete the assignment, the assignment is measuring output, not understanding.
The Equity Issue We Cannot Ignore
AI tutors do not impact all students in the same way.
Students who already have strong learning habits tend to benefit the most.
Students who struggle with motivation, reading comprehension, or executive functioning often disengage or misuse the tool.
Access is not the barrier anymore.
Effective use is.
Without intentional support, AI risks widening gaps rather than closing them.
A Necessary Counterpoint
Some will argue that students simply need more time with AI tutors.
That with continued exposure, they will learn to use them effectively.
Time alone is not enough.
Without explicit instruction and intentional design, students tend to reinforce the same patterns. They look for faster answers instead of deeper understanding.
More access does not automatically lead to better learning.
The Money Behind the AI Tutoring Push
Across the country, districts are spending heavily on tutoring.
Much of that funding accelerated during pandemic recovery, with a focus on closing learning gaps quickly. High-dosage tutoring became a priority, and vendors moved quickly to meet that demand.
Now AI tutoring tools are entering that same space.
The question is no longer whether we can provide tutoring at scale.
The question is whether the tutoring we are funding actually leads to learning.
Because the research is clear on one point.
Effective tutoring is not just access to help. It depends on:
Consistent human interaction
Strong instructional alignment
Ongoing feedback tied to student thinking
AI can support parts of that.
It does not replace it.
If districts are investing in AI tutoring as a cost-saving measure or a scalable substitute, we need to be more precise in our expectations.
What are students actually doing during that time?
Are they engaging in meaningful practice, or completing tasks more efficiently?
Are we measuring improvement in understanding, or just completion rates?
Those are different outcomes.
And they require different tools.
If we are going to invest in tutoring at scale, we need to be just as serious about how that tutoring works as we are about how much we are spending.
So Did AI Tutors Fail?
No.
But the idea behind them did.
The belief that access to an AI tutor would naturally lead to better learning outcomes is breaking down.
Because learning is not just about access to help.
It is about how that help is used.
The Line We Need to Be Honest About
We did not overestimate AI.
We underestimated how much learning depends on human behavior.
Why This Matters Right Now
This is not just about one tool.
It is about a broader assumption that keeps showing up in education technology.
If we build something powerful enough, students will use it well.
That assumption has never held up.
AI just makes the gap more visible.
A Note on Data and Privacy
AI tutors rely on student interaction data to function.
Schools need to be asking clear questions about:
What data is being collected
How it is stored
Who has access
Not all tools meet the same standards, and this cannot be an afterthought.
Preview: What Comes Next (Paid Section)
If AI tutors are not the solution, then what is?
What should independent work look like now?
What does practice look like in an AI-rich classroom?
And how do we design learning that cannot be outsourced to a tool?
The answers are already emerging. But they require a shift many schools have not fully made yet.
