Sal Khan: Khanmigo's First Chapter Changed How I Think About AI

By Sal Khan, Founder and CEO of Khan Academy

Setting the record straight

I started Khan Academy as a nonprofit because I believed that access to a great education should not depend on where you live, how much money your family has, or whether you happen to find the right tutor at the right moment.

So when I talk candidly about what we are learning from AI, it is not because I am trying to lower expectations or walk away from the work. It is the opposite. If we are serious about helping students learn, especially students who are not already well-served, then we have to be honest about what is working, what is not working yet, and what needs to get better.

Some people heard my recent comments about Khanmigo and interpreted them as “AI tutoring failed.” That is not what I said, and it is not what I believe.

What I said was that the first version of Khanmigo that we launched three years ago did not change student learning as much as many of us hoped it would. That matters. But it is also exactly the kind of thing you learn when you bring a new tool into real classrooms with real students and teachers.

I want to set the record straight and clarify three things:

1. Khan Academy works. Decades of independent data prove our platform drives substantial learning gains. My comment was a candid observation about what we learned when a brand-new medium was layered on to the existing, effective platform three years ago.

2. We pride ourselves on being transparent and honest. The edtech landscape is currently flooded with flashy, unproven AI tools that make grand promises while offering little more than wrappers around major large language models with limited safety guardrails. Khanmigo was, and remains, a pedagogically sound AI tutor with guardrails that was built for students in classrooms. Most importantly, it is anchored on Khan Academy’s demonstrably effective, human-crafted instructional and practice content.

3. The comment was in reference to the first version of our AI integration, and we used that honest reflection to build something better. We didn’t give up; we iterated. We have used those early classroom insights to build a far more effective, next-generation learning experience within our proven learning platform.

Why integration matters

The combination of Khan Academy’s world-class, human-crafted, instructional materials and Khanmigo’s ability to drive deeper engagement fixes the fundamental flaws of both traditional edtech and AI-only tools.

1. It replaces “cognitive offloading” with “cognitive onloading.”
Because Khanmigo is built directly into Khan Academy’s practice problems, it can see the problem a student is working on and discuss it with them. Instead of handing over the answer, Khanmigo shows them where they might make mistakes and provides support for arriving at the correct answer. Using data to decide which changes to Khanmigo actually engage students more in learning, we are working on new ways to help students engage in this conversation.

2. We chose integration over isolation.
In our next-generation tools, Khanmigo is no longer an add-on that requires students to recognize when they need help and understand how to ask for it. Khanmigo now prompts a student to explain how they arrived at an answer and explain their thinking. Students retain agency over whether they engage or not, but we have reduced the metacognitive load required for them to do so. This is not AI for AI’s sake. It is leveraging AI to increase students’ learning and reflection.

3. It multiplies teacher capacity.
A teacher with 30 students cannot sit next to every child to catch a mistake the moment it happens. Khan Academy has always provided teachers with a data-driven map of who has mastered a concept and who is stuck. Now, Khanmigo acts as the on-the-ground coach. It handles the immediate, granular feedback, which frees up the teacher to do the deeply human work of motivating a discouraged student or running a small-group intervention.

What we’re seeing in schools

We aren’t incorporating AI for AI’s sake. We are using it to provide scaffolding and feedback during independent practice.

Let’s be absolutely clear about what our data actually showed. For students who engaged, Khanmigo did what we designed it to do. It probed reasoning, supported students as they worked through a concept, pushed back on logical fallacies, and behaved like a human tutor.

But we also learned something: if Khanmigo was going to help students think harder, it had to become more central to the practice experience. The AI could not just sit next to the content. It had to be woven into it. We had to make productive struggle harder to sidestep.

Unlike many AI tools that are still making broad promises, Khan Academy’s work is grounded in evidence. In New Jersey, students in Newark Public Schools using Khan Academy saw math gains on their state assessments.

That matters because Newark is not a boutique experiment. It is a real district serving real students. And it shows what is possible when trusted content, teacher insight, and thoughtful AI work together.

Expectations and reality

When I first saw the early versions of GPT models in August of 2022, I was amazed by what they could already do and was excited by the potential. Some of that potential has been realized. The models are far better at math now than they were at the time.

However, other aspects of that potential have not been realized. AI is not good at generating standards-aligned content on the fly, for example. We have learned that it is far better when paired with our expert human-created content. In May 2024, I made a video with my son that showed a demo of a tutor that could see what we were drawing and converse back and forth. We are only now starting to see some of the model capabilities that would allow Khanmigo to interact with visuals outside of a demo environment.

Some things I envisioned will happen more slowly. Some may not work out. And some will go on to help improve learner outcomes. We will keep being transparent with you about what works, what doesn’t, and what we’re learning along the way.

The path forward: scaffolds, not shortcuts

The current debate in education is gridlocked in a false dichotomy: sweeping bans on technology versus a tech-utopian free-for-all. Both sides are wrong. The answer isn’t to take away the laptops; the answer is to give students tools that actually support student learning.

Families with time and resources will always find ways to provide elite tutoring for their children at home. If we respond to the challenges of AI by banning it in schools, we are putting up a velvet rope disguised as caution tape, denying under-resourced students the very tools that can level the playing field.

The combination of Khan Academy and Khanmigo is a strong solution because it strikes the right balance. It provides the strict guardrails, safety protocols, and teacher oversight that schools require while also delivering the personalized, rigorous intellectual push that students need.