A Conversation with Sal Khan

Type: Article
Topics: School Administrator Magazine, Technology & AI

October 01, 2025

The learning innovator on AI’s potentials and perils, the efficiencies it will bring to schooling and its use in places with limited financing
Sal Khan, founder of Khan Academy in Mountain View, Calif., has created a tool that uses artificial intelligence to personalize tutoring support for students. PHOTO BY RACHEL MURRAY/GETTY IMAGES

Sal Khan believes his mission is to “provide a free, world-class education to anyone, anywhere” through his Khan Academy as well as his newly derived artificial intelligence-powered personalized tutoring tool and teaching assistant, Khanmigo.

The latter, introduced in March 2023, expects to reach one million elementary and secondary school students in the U.S. this school year, and its adoption is growing rapidly.

School Administrator magazine asked Khan to discuss his work with education leaders on his AI-infused education platform and his book Brave New Words: How AI Will Revolutionize Education (and why that is a good thing). (Khan was a keynote speaker at AASA’s 2018 National Conference on Education, and his company’s chief learning officer, Kristen DiCerbo, addressed AASA’s AI Super Summit 1.0 last January. See her article, right.)

The interviewer, Howard Carlson, is executive director of the Greater Phoenix Educational Management Council and the Arizona Educational Management Council. Carlson is the author of two books, So Now You’re the Superintendent and Accelerated Wisdom: 50 Practical Insights for Today’s Superintendent.

The Q&A that follows has been edited for clarity and length.

Carlson: In your book, Brave New Words: How AI will Revolutionize Education, you speak to working with OpenAI to build an “AI-infused education platform” called Khanmigo. What do you want superintendents to understand regarding how Khanmigo will “revolutionize” education?

Khan: When ChatGPT first came out, it was seen as a cheating tool. But I think very quickly educators realized that this technology was here to stay. If someone were to give AI a framework to be used safely and thoughtfully, it could advance some existing goals around pedagogy, traditional learning outcomes and future skills.

When we saw what the capabilities were about three years ago, we said, “We just need to put in place the right guardrails.” This cannot be a cheating tool, it must be Socratic, it needs to be transparent for educators.

We also realized that AI should not be viewed in a vacuum. AI is not going to be a panacea by itself, but rather how it will be used in conjunction with other things. All the efficacy studies have shown that when students engage in practice at their level and get feedback, real learning occurs.

We know teachers must be freed from some non-teaching parts of their job. If they are spending 10-15 hours per week on lesson planning and progress report writing, AI can free up a teacher’s time for other more meaningful activities. Administrators have had few levers to directly support and understand what is occurring in the classroom. AI, in conjunction with Khan Academy, enables district administrators to set goals with teachers and students to get alignment within the system.

Carlson: What are your particular concerns about AI-fueled cheating?

Khan: Often policies differ from teacher to teacher, and I have spoken to many about the importance of ensuring everyone is on the same page. Tools that claim they can detect AI writing are snake oil. They have 40 to 50 percent false positive rates. If used, you are going to accuse a lot of honest kids of cheating. The best thing is to give clear guidance regarding how students can, and in which situations they should, use AI.

We also should be doing more in-class writing. We have been developing a program called Writing Coach that helps older students write a term paper. The AI does not do it for them but rather supports them like a tutor might do, so it makes the process transparent to the teacher.

Carlson: You launched Khanmigo in March 2023 and by 2024 you were piloting the application with over 30,000 teachers and students. What early results are you seeing regarding the use of the AI tutor, and what has surprised you thus far?

Khan: We expect to reach one million this year (up from 700,000 last year). Khanmigo reinforces the idea that AI should not be seen in a vacuum. Twenty percent of students will use Khanmigo and say, “I can get a lot out of this,” but a lot of students do not know how to get started, and this is not an AI issue. Teachers tell us these same students have challenges engaging in the classroom too.

This is where we are investing more of our time. We are designing Khanmigo to be both the teacher’s and the student’s guide, and this is its draw. It is more than just a tutor. It brings the content they need to them and is constantly communicating with the teacher. It’s surfacing actionable things for the teacher to do.

For example, Khanmigo will suggest “click here and assign a tutoring session with this student.” This information also can be used in the preparation phase as the teacher is planning lessons.

Carlson: What’s your vision for Khanmigo? Does it become interactive like Jarvis in the “Iron Man” movies, or do you see other possibilities?

Khan: I start by thinking about what educators might want if they had more resources. They would say they want teaching assistants who could help them do lesson plans, grade papers, complete progress reports or tutor their students and always be available and in constant communication with them so they could serve their students better.

In the classroom, AI can play a role in helping solve each of these challenges. For a district administrator, the advantage is being able to know what is going on, being able to pick curricula and provide professional development. District administrators often feel like they don’t know what is going on until they see the test scores at the end of the year.

As leader of a nonprofit, Sal Khan says he intends to bring down the cost for school districts using his tools, which he says currently run $10 to $15 per student per year. PHOTO COURTESY OF KHAN ACADEMY

These areas are problems to solve, not just with AI, but with other technological tools, human systems and professional development. There are some school systems, such as Newark, N.J., that are doing some incredible things on this front every week. The head of the math department can look at a dashboard and see where there is engagement and where there isn’t. He can go down to the individual student level in real time, or down to a class or school level and see not only engagement, but also which skills are being pursued to mastery. This is empowering for an administrator, and unless you have the right tools, this is not possible.

Carlson: As I think about school districts across the country facing financial challenges, how do you believe Khanmigo can be more broadly implemented in an environment dictated by limited resources?

Khan: Seven or eight years ago, we started going to school districts and saying, “There are these efficacy studies on Khan Academy, and you have teachers already using the application. Why don’t you begin using us more systemically?”

Pretty much every superintendent or chief academic officer said to us, “Yes, we are aware of the application’s success.” But they said, “if you really want us to use you more broadly, we need more support, more training, district-level dashboards and integration with our rostering systems.”

That is when we started the journey toward developing the Khan Academy Districts offerings, which Khanmigo is now a component of today.

Our goal has always been to make this more affordable. In fact, we do not even charge for our product development, research and development, which are funded through philanthropy as a nonprofit. We only charge the districts our marginal costs, which is how we get down to this $10 to $15 per student per year. And I know that even a nickel is often not available for most school districts, but if we address these core issues around student outcomes, we can help them be more efficient.

I know districts need to spend a lot on credit recovery so Khan Academy could be a powerful tool in this space. If we drive engagement among students and accelerate learning by 20 or 30 percent in every subject from preK through college, that is going to save a ton of money.

If teachers can reduce planning time from 15-20 hours per week to 2-3 hours per week, you are going to enhance retention, which saves money. If administrators can use the tool when teachers exit mid-year or are on leave, the tools can keep learning momentum intact.

Carlson: So you are talking about creating efficiencies to free up funding. At this point, are you researching the types of efficiencies that districts are achieving by using Khanmigo?

Khan: That is a good idea. I am going to have to tell my team to make some case studies like this.

Yes, we see that, if students get to a certain level of engagement, they will grow 30, 40 or 50 percent more than expected. The hard part is how do you get all kids to engage.

What we see in Newark is that when there is alignment with district leadership and teachers, student engagement increases. In Newark, over 70 percent of students are engaging with Khan Academy at a level we know correlates with real academic gain – 30 minutes per week.

I suspect Khanmigo is reducing the number of Newark students in credit recovery and the number of students who require additional support services. I have heard universally that 25 percent of algebra students require credit recovery with some vendors charging districts $300 per student per credit for these services. If we can reduce the number of students needing credit recovery, it more than pays for itself.

Carlson: Not all students will prompt Khanmigo with well-written questions, which can limit AI’s ability to provide quality information. How does Khanmigo address this reality? Are there downsides or dangers from careless use of the tool?

Khan: I will take it in reverse order. Obviously, students saying anything to AI introduces risk and this is where we have done a lot of work regarding transparency for teachers and administrators. AI in isolation is of limited benefit. AI in conjunction with vetted content that has been proven to be effective is where you start to see results.

If we just wait for students to ask the right questions, you will never get the benefit.

That is why they need to be on a platform where the teacher and the platform are directing them to work on and practice things that will be good for them. The AI is there and can recognize when they need help and jump in to encourage engagement and understanding.

The AI tool should not be like ChatGPT, just waiting for someone to ask the question. It should be proactive, it should be integrated with an efficacious learning platform, it should be standards aligned, and it should include safety measures.

Administrators need to think deeply about the problems they are trying to solve, not just jumping on the AI bandwagon. Be skeptical if it is AI in isolation. It should be AI that is augmenting things that have already been proven to drive outcomes or save teachers’ time.

Carlson: Based on the feedback you have received from participating school districts to date, how challenging is it for schools to implement Khanmigo?

Khan: When the initiative is a priority for school and district leaders, it is going well.

Newark, as I mentioned, is one of the best cases, but many districts have pulled it off. That said, we know there are things we need to do to make it easier for district leaders to align their educators on how to use these tools. And that’s professional development, which we do, but also making the tool self-explanatory.

Carlson: As generative AI continues to develop, how will the teaching profession and K-12 education in general change?

Khan: The teaching profession is going to move up the value chain. Over time, teachers are going to spend less time on lesson planning, report writing, progress reports, grading papers and will be able to place that found time on providing rich, interactive lessons that are more personalized to not only meet the interests of the students, but also the needs of the students.

Teachers will have much better data and will have access to an assistant that can help one make sense of what they are seeing and help them to take action. So I think generative AI makes a teacher’s life much more pleasant.

If you tell teachers to imagine a world where you have three or four amazing teaching assistants who are tireless, energetic and creative, and then ask them — “how would this change your job?” — most teachers would say, “that would be amazing! I would be able to do more for my students.” So, I think any job that has a human-leaning aspect will become more important because many current tasks take away from focusing on the human aspect of the job.

Carlson: As a final word, if you were a school superintendent, how would you navigate this moment in time related to implementation of AI in our K-12 system? What would you be thinking and doing? How would you be preparing and acting? Also, what policies should school boards be pursuing regarding AI in the coming year?

Khan: I will start with the policy component. One is an AI cheating policy. Making clear that there are no tools where you can accurately detect AI. Instead, use AI tools that focus on transparency in the writing process, like Khan Academy Writing Coach, or even Google Doc plugins that show what a student has accomplished. Also, outline how AI can be used in assignments at different grade levels.

If I am a superintendent or a chief academic officer, I am thinking I do not want to be a late adopter, but I also want to be careful trying new things. I would focus on core items I am trying to solve.

As a superintendent, one of the top three things you want to achieve is to increase student outcomes. Make that your true north, not just using AI, but rather asking yourself “how do I use AI to improve student success?” You can then look for things, not just AI tools, which show good evidence of achieving results. Yes, AI can be part of that mix, but I would be very skeptical of using AI in isolation. 

Howard Carlson

Executive Director

Greater Phoenix Educational Management Council and the Arizona Educational Management Council

The Possibilities of Personalized Learning via Artificial Intelligence
Kristen DiCerbo. PHOTO COURTESY OF KHAN ACADEMY

By Kristen DiCerbo

While artificial intelligence is rightfully the shiny new technology everyone is talking about, the problems it might address are not new. In classrooms, we long have faced the challenge of giving all students the support they need when students are working at many different levels of achievement and motivation.

Students need to practice new skills they are learning and apply them in different contexts to gain the fluency needed and to embed the new skills in long-term memory. Lev Vygotsky, a cognitive psychologist from a century ago, theorized that learning occurs in the gap between what students can do independently and what they can do with the support of a more capable person. If students are at many different levels, it is nearly impossible for a single classroom teacher to identify that spot for each student and provide just the correct level of support.

We also know that receiving feedback is key to learning. In a review of the research on feedback, Valerie Shute, an educational psychologist at Florida State University, found consistent evidence that feedback offering explanations for why something is correct or incorrect (“elaborated feedback”) results in more learning than feedback that only tells the student if they are right or wrong (“correctness feedback”).

In addition, for students learning something new, feedback that immediately follows a student’s response leads to more learning than delayed feedback. Again, it’s a challenge for one teacher to provide immediate, elaborated feedback. In the case of writing assignments, teachers spend evenings and weekends grading a single assignment and students still don’t get the feedback until days later, by which point most of the teacher’s detailed comments often are given less attention than is desired for learning.

Feedback Illustrations

If we define personalized learning based on findings about how people learn, we can see how well-designed, education-specific applications of new generative AI technology might personalize learning in a way that helps students learn.

The promise of generative AI for personalization lies in the ability to provide support and immediate feedback tailored to individual students as they develop new knowledge and skills. Here are a few examples of how students are able to benefit from personalized support and feedback with generative AI tools designed for learning.

Timely translations: Generative AI is good at interacting in common languages spoken by students. When learning new concepts and ideas, students are able to get explanations of them in their home language when English explanations are not making sense and there is no one else to help translate.

Real-time writing support. Automated essay scoring streamlined the task of grading in the days before generative AI, but it never could provide specific feedback and advice on individual essays. Now, students struggling to write a persuasive essay can get feedback in the midst of writing a draft on whether the essay adequately supports the argument with evidence. Students can get examples of good use of evidence, make writing changes and check to see whether they have improved the argument.

Individualized math assistance. When solving an algebra problem, a student may struggle with least-common multiples. If the student is working with a generative AI math tool, when getting to that step, it will provide the support needed to apply the prerequisite skill to this new situation. In a classroom, if stuck on this step, the student would likely be sitting with a hand raised waiting for the teacher, unable to progress.

Student Preferences

Apart from these individual tasks, educational platforms that incorporate AI along with high-quality instructional materials are able to establish student profiles. Khan Academy’s Khanmigo has estimates of students’ levels for skills they have practiced, can store information about their preferred reading level for AI responses and their interests and has memory of previous conversations. As a result, students’ practice sessions throughout the year can target the levels of support and feedback appropriate for them.

This type of personalization with AI will not happen well with tools built for general use. Only AI tools built for educational purposes can provide this type of support. Significant work has to be done to provide instructions to AI models to act in pedagogically sound ways and to do it consistently.

However, with careful guidance and instruction, students can receive more tailored support and practice than they have ever been able to before, and that has great potential to improve learning outcomes.

Kristen DiCerbo is chief learning officer of Khan Academy in Mountain View, Calif. 

Advertisement

Advertisement


Advertisement

Advertisement