The Promise We Thought We Had
When Khan Academy launched Khanmigo back in 2023, the education world held its breath. Here, finally, was an AI tutor that could theoretically give every student their own personal guide—someone patient, infinitely available, and armed with pedagogical precision. The numbers initially looked good. By late 2025, the platform had logged over 50 million tutoring sessions, and middle school math completion rates jumped 23 percent among users. For Title I schools especially, where budget constraints make hiring additional tutors impossible, the Gates Foundation’s Q3 2025 commitment of $15 million to expand Khanmigo across 12 states felt like a genuine breakthrough. This wasn’t just another edtech promise. This felt real.
But here’s what I’ve learned after two decades in classrooms: promising numbers and actual learning are not always the same thing. They’re not even close sometimes. So I spent the last few months genuinely interrogating what Khanmigo 2.0 does well and where it stumbles in ways that matter for how your brain actually learns.
Where Khanmigo Actually Nails It: The Socratic Approach
The part that made me sit up and take notice involves how Khanmigo 2.0 handles questioning. Recent research from Stanford PACE Center AI in Education Reports found something remarkably clear: AI tutors that ask Socratic questions instead of offering direct answers perform 31 percent better on retention tests two weeks later. Thirty-one percent. That’s the difference between a student genuinely understanding photosynthesis and a student who memorized the steps for a quiz.
Khanmigo 2.0 has internalized this. When a student struggles with, say, setting up a multi-step algebra problem, the system doesn’t just say “multiply both sides by three.” It asks: What do you notice about the equation? What would happen if you moved all the variables to one side? What operation undoes multiplication? This mirrors what actually happens in my classroom when I’m helping a student think through something rather than handing them the answer. The research backs this up completely, and watching Khanmigo handle this correctly feels like vindication that learning science finally made it into mainstream AI design.
The platform’s work with English language learners deserves special mention because it’s where the gains feel most meaningful. According to Khan Academy’s own 2025 efficacy report, ELL students showed a 41 percent improvement in reading comprehension scores after eight weeks of Khanmigo use. That’s not a rounding error. That matters profoundly for students whose learning barriers have nothing to do with ability and everything to do with access to patient, culturally responsive instruction.
The Dangerous Problem: Students Stop Struggling Productively
But here’s where I have to be honest about something that concerns me deeply. A January 2026 RAND Corporation survey found that 67 percent of teachers using AI tutoring tools reported that their students became less likely to struggle productively with difficult problems. Let me translate what that means in learning science terms: struggle is where learning actually lives. When you wrestle with a challenging concept, your brain forms stronger neural pathways than when someone explains it to you. Productive struggle—the kind that feels hard but achievable—is essential to real understanding.
The risk with an AI tutor that’s always available and always helpful is that it can short-circuit this process. A student gets stuck, the AI jumps in with a guiding question, the student finds the answer, and everyone feels good. But the student’s brain never learned to sit with confusion. They never developed the metacognitive skills that come from pushing through difficulty on their own first. This is subtle enough that it’s easy to miss, especially when completion rates are climbing and assessment scores are improving. But it’s the difference between a student who can solve a problem and a student who is becoming a problem-solver.
Think about it this way: if I let every student check with me the moment they felt uncertain, would my classroom improve or would it create learned helplessness? That tension doesn’t disappear just because the helper is artificial. It gets more complicated.
What Khanmigo 2.0 Gets Wrong: Relationship and Accountability
There’s something about the human relationship between a teacher and student that AI, no matter how sophisticated, cannot replicate. When I remember that you found fractions impossible in September but mastered them by April, I’m not just pulling up data points. I’m acknowledging your growth. I’m communicating that I see you. That I believed you could do this. I have accountability to you in a way that an algorithm doesn’t.
Khanmigo 2.0 is extraordinarily good at being consistent, patient, and evidence-based. But it has no stake in your success beyond the parameters of its design. It cannot call your parent because it notices you’re slipping. It cannot catch the moment when you’re overwhelmed versus when you’re just procrastinating. It cannot adjust its teaching style when it senses that you learn through humor or movement or talking things out loud. These aren’t flaws in Khanmigo specifically. They’re inherent limitations of the technology, and pretending otherwise would be dishonest.
The research showing gains in math completion rates and reading comprehension is genuinely impressive. But those metrics don’t capture everything that matters in learning. They don’t measure whether a student developed genuine curiosity about the subject. They don’t track whether a student learned to advocate for themselves. They don’t capture the quiet moment when a struggling student finally believed they could do hard things.
How to Use Khanmigo Wisely (Spoiler: It’s Supplement, Not Substitute)
Here’s what I actually recommend if you’re considering Khanmigo 2.0, whether you’re a teacher, parent, or student. Use it strategically as a supplement, not a replacement. It excels at providing immediate, responsive help with discrete skills and at giving students practice with Socratic questioning, which genuinely deepens retention. The platform is particularly strong for students who are shy about asking questions in class or who need explanations at 11 p.m. on a Sunday. For ELL students and others who benefit from multiple exposures to content, the research is clear that it works.
But preserve productive struggle. Maybe Khanmigo handles practice problems, but you or your teacher handles the initial problem-solving attempt without support. Maybe you use it to check your understanding after you’ve genuinely tried. Maybe you save it for reinforcement rather than first exposure. Build in moments where the answer is not available, where confusion is the actual goal.
Check out the detailed research yourself on Khan Academy Khanmigo Research and Efficacy so you understand what the data actually supports. The research is solid where it’s solid, and it’s worth engaging with on its own terms rather than taking anyone’s word for it—including mine.
What’s your experience been with AI tutoring tools? I’m genuinely curious whether these concerns ring true in your classroom or home school, or whether you’ve found ways to use these tools that navigate these tensions better than I’m imagining.
