Khan Academy’s Khanmigo After One Year: What the Data Actually Shows About AI Tutoring in Schools
The Scale Question: From Pilot to Real Classrooms
When Khan Academy launched Khanmigo in limited beta back in 2023, it reached about 200,000 students. By the end of 2025, that number had climbed to 1.5 million users across 130 countries. That’s not just growth. That’s the moment when an experiment stops being a curiosity and becomes infrastructure that real teachers and students are actually depending on.
But here’s what matters more than the headline number: those students are in actual classrooms with actual teachers who have actual deadlines and grading piles and students who hate fractions. The question isn’t whether an AI tutor works in theory. It’s whether it works when a seventh grader uses it at 10 p.m. the night before an algebra test, or when a teacher with 150 students needs to figure out who’s ready to move forward.
Math Works. Writing Doesn’t. Here’s Why That Matters.
The WestEd randomized controlled trial across 47 schools in 2025 gave us the clearest evidence yet. Students using Khanmigo for math showed a statistically significant 0.15 standard deviation improvement in algebra readiness scores after one semester. That’s real. It’s not enormous, but it’s consistent, it’s measurable, and it’s there.
Then they tested the essay coaching feature. The result? Nothing. No statistically significant improvement in writing skills. On the surface, that sounds like a failure. But the researchers offered something more useful than disappointment: a hypothesis. The tool encouraged revision over generative thinking. Students got trapped in the loop of “fix this comma, improve this transition” without wrestling with the harder work of building an argument from nothing.
This distinction matters enormously. Math problems have clearer structure. You can get stuck on a specific operation or concept, and a well-timed hint that doesn’t give away the answer can unlock your thinking. Writing is messier. The hard part often isn’t execution. It’s deciding what you actually want to say. An AI that helps you polish what you’ve already decided to write might actually get in the way of the cognitive work that makes you a better writer.
Teachers Are Using It. And Not for What You’d Expect.
Here’s something that surprised me: teachers using the Khan Academy Khanmigo for Teachers dashboard spent an average of 37 minutes less per week on progress monitoring and administrative paperwork. That’s based on time-use surveys from 3,200 participating teachers. For a profession where paperwork often feels like the main job and actual teaching is what you squeeze in between, 37 minutes per week is real time back.
But here’s what this actually means: teachers aren’t using Khanmigo primarily as a tutoring tool for students. They’re using it to automate the grading and progress tracking that was already taking them hours. The AI handles the administrative overhead, which frees teachers to do what they’re actually supposed to do: give feedback that matters, notice patterns in how their students think, and adjust their teaching in real time.
That’s not revolutionary. But it’s valuable. And it’s different from the “AI replaces the teacher” narrative that gets attention at conferences.
Why Students Trust Hints More Than Answers
Research from the Christensen Institute EdTech Research in 2025 surveyed students about their homework help preferences. Sixty-eight percent preferred Khanmigo’s Socratic approach (hints, not direct answers) over ChatGPT for homework help. The reason they gave? Reduced anxiety about cheating.
This is where learning science catches up to what good teachers have always known: students learn better when they do the cognitive work themselves. But more than that, they feel better about it. There’s a real psychological difference between “the AI gave me the answer” and “I figured it out with help.” One feels like cheating. The other feels like learning. That distinction shapes whether students actually retain what they work through.
The hint-based approach also maps onto what we know about productive struggle. When a student gets stuck and then receives a well-calibrated hint pointing them in the right direction without handing them the solution, they’re doing the work. Their brain is engaged in retrieval, reasoning, and problem-solving. That’s the cognitive workout that builds understanding.
What This Data Means for Your Classroom (or Your Kid’s)
After one year of real classroom data, here’s what we actually know: Khanmigo works for structured, step-based learning in math. It helps teachers with administrative overhead. And students engage with it differently than they engage with unrestricted AI tools because the design of the tool makes them feel like learning partners instead of answer machines.
It doesn’t work for everything. Writing instruction is more complex than the tool currently handles. The improvement in math is meaningful but modest. It’s not a replacement for skilled teaching. It’s a tool that does specific things well and stays honest about what it doesn’t do well.
If you’re a teacher wondering whether to try it: think about where your administrative burden is highest and where your students get stuck on predictable, diagnosable problems. If you’re a parent wondering whether to let your kid use it: watch whether it’s actually making your kid work, or whether it’s letting them passively consume solutions. If you’re a student: use it the way it’s designed to be used. Get the hint. Figure out the next step. That’s where learning lives.
The data after one year is telling us something important: AI tutoring isn’t magic. But when it’s designed around what actually helps people learn, and measured honestly against what it can and cannot do, it becomes genuinely useful. That’s worth paying attention to.
Comments are Disabled