The Math Works. Really Works.

Let me start with what’s genuinely impressive because I think we owe it to the people building this stuff to acknowledge when they nail something. Khan Academy released updated efficacy data in 2025, and the middle school math completion rates jumped 23 percent among users who engaged regularly with Khanmigo. That’s not a marginal improvement. That’s the kind of number that makes you sit up and pay attention.

But here’s what matters more than the headline stat: English language learners showed a 41 percent improvement in reading comprehension scores after eight weeks of using the platform. I’ve taught enough students with interrupted formal education and language barriers to know how stubborn those reading comprehension gaps can be. A 41 percent bump in eight weeks? That’s real. That’s the kind of targeted support that works because it meets students exactly where their confusion lives, then builds from there.

The platform has now logged over 50 million tutoring sessions since launching in 2023. That volume matters because it means the system has seen every version of “I don’t understand fractions” that exists. It’s learned from millions of moments when a student went from stuck to unstuck.

Why the Socratic Method Actually Matters (And Why Most AI Misses It)

Here’s where I get genuinely excited, and also where I see the cracks forming. Research from Stanford’s Graduate School of Education in 2025 found something crucial: AI tutors that ask Socratic questions outperform those that hand you the answer by 31 percent on retention tests two weeks later. Thirty-one percent. That’s enormous for long-term learning.

The difference is simple to explain but hard to execute. When I tutor you through a calculus proof by asking “what do you notice about the slope here?” versus telling you “the slope is negative,” something different happens in your brain. The first version requires you to do cognitive work. It’s uncomfortable. It’s also how actual learning happens. You’re building a mental model, not just absorbing information.

Khanmigo 2.0 does this better than its predecessor. The system actually resists the urge to immediately solve your problem. It asks follow-up questions. It lets you sit with the discomfort for a beat. Khan Academy Khanmigo Research and Efficacy reports show they’ve deliberately engineered this behavior into the platform. That’s not accidental. That’s a design choice made by people who understand how learning actually works.

The Equity Win That Matters Most

The Gates Foundation committed 15 million dollars in Q3 2025 to expand Khanmigo access to Title I schools across a dozen states. Translation: schools serving the students with the fewest resources, the most interrupted schooling, the least access to private tutoring, are now getting an AI tutor available 24/7.

This isn’t a small thing. I’ve worked in schools where families couldn’t afford the 60-dollar-an-hour tutoring center down the street. Where a student’s learning gap didn’t get addressed because their family was choosing between tutoring and rent. The access problem is real. An AI tutor that’s consistently available, never tired, never impatient, and actually effective changes the equation considerably.

The growth in ELL student performance I mentioned earlier sits inside this equity story. Language learners often need different kinds of scaffolding, more repetition, different pacing. They need a tutor who isn’t going to get frustrated or make them feel rushed. An AI system can provide that endlessly.

Where It Falls Apart: The Productive Struggle Problem

Now for the harder conversation. A January 2026 RAND Corporation survey asked teachers about their experience with AI tutoring tools, and 67 percent reported something troubling: their students became less likely to engage in productive struggle with difficult problems. They wanted the AI to just solve it for them faster.

This is the ghost in the machine. You can design the most elegant Socratic questioning system imaginable, but if a student discovers they can game the system by asking slightly different questions until someone just tells them the answer, they will. The friction that makes learning work? Students hate that friction when they’re tired or overwhelmed or already frustrated.

I see this happen in real classrooms. A student works with Khanmigo, makes progress, then encounters a textbook problem that requires that struggle-mode thinking. They freeze. They’ve become dependent on the scaffolding without internalizing the actual problem-solving process. The platform solved one problem (access to tutoring) while creating a new one (atrophy of tolerance for difficulty).

What Actually Matters When You’re Choosing

If you’re a teacher or administrator deciding whether to implement this stuff, here’s what I’d tell you: Khanmigo 2.0 is a genuinely useful tool for specific populations in specific contexts. The data on ELL students and struggling middle school math learners is solid. The Socratic questioning approach works when students actually engage with it.

But it’s not a replacement for knowing your students. It’s not a substitute for a teacher who understands that some kids need pressure and some need permission to take their time. It’s a supplement. A really good supplement. Stanford PACE Center AI in Education Reports are showing us that AI works best when it enhances human judgment, not when it tries to replace it.

The real question isn’t whether Khanmigo works. It’s whether your students know how to work. Whether they can sit with a hard problem without immediately reaching for the answer. Whether they’ve practiced failing in safe ways so they’re not terrified of failure in real ways. Those things are still on us. The AI can help. It can’t do that part.

What’s your experience been? Are you using AI tutoring tools in your classroom or with your own learner? I’d genuinely love to hear what you’re seeing on the ground.