Sal Khan: On the Frontlines of EdTech and AI

Edtech Insiders joined the first Common Sense Summit on America's Kids and Families hosted by Common Sense Media at Pier 27 in San Francisco on January 28–30.

The conference brought together advocates, researchers, youth leaders, policymakers, and other experts to take stock of America's kids and families and explore solutions to the most pressing issues across four core topic areas: kids and technology, youth mental health, early childhood education, and K–12 education.

In this special episode, Alex and Ben interviews Sal Khan.

Sal Khan is the founder and CEO of Khan Academy, a nonprofit educational organization that offers free lessons in math, science and humanities, as well as tools for parents, teachers and districts to track student progress. Khan Academy is piloting an AI guide called Khanmigo that is a tutor and teaching assistant.

Sal Khan founded Khan Academy in 2008. Khan Academy is a nonprofit educational organization with 220 employees. In addition to math, they offer free lessons in the sciences and humanities. Outside of the Khanmigo pilot, more than 500 public school districts and schools across the United States partner with Khan Academy (up from nine before the pandemic). Khan Academy is used in more than 190 countries and is available in more than 50 languages. They have more than 160 million registered users worldwide.

Transcript

Ben Kornell:

I am super excited, super honored, to have Sal Khan- the man, the myth, the legend. You all know everything else. And we're so excited to have you join us. You started Khan Academy in 2008, and we're here in 2024. What turned out the way we expected and what surprised you?

Sal Khan:

Oh, boy, that's a big question. A lot of folks might already know that, in some ways, Khan Academy started even before 2008- it was 2004. I was a year out of business school. My original background was in tech, I had just gotten married and my family was visiting me from New Orleans. I learned my 12 year old cousin Nadia needed help with math and I offered to tutor her remotely. She agrees, and that works out well- she goes from being a struggling student to being a very strong student in math.

Word spreads in my family that free tutoring is going on. Before I know it, I'm tutoring 10-15 cousins, family, friends. Obviously, I wanted to help my family, and it seemed like I was able to help my family— but even then my interest— I even used to tell folks, including my wife at the time (and still my wife), she always used to give me a little bit of flak like “you've gotten such a good education; is working at a hedge fund really the highest and best use out of it?” And I said, “Well, you know, we get a lot of debts, between your med school debt and my grad school debt and undergrad debt (laughs)”... but I always used to tell her and my friends, “well, I want to do this hedge fund long enough so that one day I can start a school on my own terms.”

That was always in the back of my mind: that a lot of people say those types of things that have never really happened. So when my cousins needed help, I said no, this is kind of a cool way for me to start understanding the issue a little bit more than I had up to that point. And I immediately saw patterns. I saw the pattern that the reason why most of my cousins were struggling - it wasn't because they weren't bright, it wasn't because they didn't have good teachers, it wasn't because they didn't go to good schools. It's because they had gaps. I mean, now everyone calls it unfinished learning.

They had unfinished learning, it was either forgotten or a regression, or they never learned in the first place. And so I saw, especially a lot of my cousins would say pre algebra, algebra, they had weaknesses in negative numbers and fraction arithmetic, dividing decimals. So I started to write software for them, so that they can get as much practice as needed to really get fluency in those types of skills so that when they get into pre-algebra, they're not sitting there and thinking about “how do I divide these decimals?”

That was the first Khan Academy, as a way of me scaling myself as a tutor. And in 2006, a friend suggested I make it into a supplement, and then those took on a life of their own, and that's what most people associate with Khan Academy.

And in terms of surprise, you know, I have an original background in tech, and I think when I started writing software. it was obvious to me in theory, it could scale. And I was working with my cousins, but in theory, maybe one day it could be 15 thousand, 15 million. And then I would say, “Sal, stop thinking delusionally; you're just doing this for your family”. In the late 90s, I had failed startups and it was just so emotionally taxing. Like, “this is not a startup, this is a family project. Don't ask me about it.” But I did think about it.

And then what surprised me is even as a hobby, from 2004 to 2009, it was gaining traction. In 2008, as you mentioned, we incorporated as a nonprofit. I don't own Khan Academy, no one does. Admission free world class education for anyone, anywhere. Even then I thought I was gonna keep my hedge fund job; this was going to be like my side nonprofit, but in 2009 I stopped focusing on my day job and took the plunge. Then what surprised me- what was the downside to that? No one seemed to care that there might be a way of offering a lot of academic support and support for teachers in a very scalable way in a way that few nonprofits can scale.

But then eventually, but all of a sudden went from zero to a lot where folks like Ann and John Doerr and I didn't even know that Bill Gates was using Khan Academy with his own kids, executives at Google were using Khan Academy with their own kids and we started growing and I think that's what a lot of folks really started to know Khan Academy.

Since then, on one level, if you told me in 2008 that in 2024, Khan Academy would have 160 million registered users I’d be like, Oh, my God, that's incredible. But at the same time, I would have assumed that with 160 million registered users, that the education system was fundamentally different. That personalization would be the norm, that every teacher would be empowered with really good data, really actionable things, right where they need it, that most students wouldn’t just be promoted forward, even though they have this unfinished learning- these gaps. And it hasn't happened. Unfortunately, in a lot of cases, maybe because of the pandemic, we've regressed a bit.

Ben Kornell:

Yeah, so the impact of new technology, so it's actually a critical platform in your tutoring taking off at scale. And now here we are in 2024 and it's actually not a new muscle you're flexing at Khan Academy. Whatever the wave is, where kids are, and where kids can benefit, you've doubled down on that. Tell us a little bit about how this has come into your thinking. And I know it wasn't an overnight decision that we’re doing this- we're doing a journey. How did that come about?

Sal Khan:

Yeah. And you know, you know, if I talk about AI, when I just talked about in terms of my journey as a tutor, if you think about everything that Khan Academy’s been doing- from 2004 with Nadia, to 2005, writing software, 2006, making videos, becoming a real non-profit, everything we've been doing is actually trying to scale the type of personalization that you could do as a tutor. And science fiction has given us a lot of examples. The Diamond Age: The Young Lady's Illustrated Primer showed an AI that can help tutor you. In Ender's Game, the AI is helping tutor these kids who are going to save humanity from an existential crisis. So that's always in the back of our minds. But how can we make this better and better for personalization?

Once again, we've never viewed it as any way a replacement for human beings, we said, hey, it could raise the floor. If you're not, you're in rural Alaska, or you're in a village in India, and you have nothing, we want to raise the floor. But the ideal is, you have caring human beings in your life. And we want to arm you and them to help you learn more. So that's always been the key.

So we've always had this view of like, okay, we will kind of want to be a tutor for every student. And we always want to be a teaching assistant for every teacher. Then summer 2022, Sam Altman, the CEO of OpenAI, and Greg Brockman the President, shot me an email, and they said, “Hey, we're about to finish training our next generation model”— which would end up being GPT4— this is them saying “we think this is the model that's going to get people excited, but maybe even a little scared, because it's going to be so powerful. So we want to launch with social positive use cases, with organizations that people trust that have the technical capability to take advantage of this.

And this is their own words, they said “The only thing we can think of is Khan Academy.” And so I took the meeting. I knew of them. I knew that GPT 1,2, 3 was super cool from a technology point of view, but I didn't think it was ready for primetime. But when they showed what was possible… actually there's a little bit of a backstory.

When GPT3 came out, Bill Gates had apparently challenged them. He said, This is cute, but I’ll only be impressed if this can pass an AP biology exam. And so they had that in the back of their mind and we have a lot of AP biology classes, to see whether they've reached their bar.

But when they showed us GPT4, I think, myself and Kristen, who's the Chief Learning Officer, we were probably one of the first 20 people in the world, even before a lot of OpenAI saw it. Even though it had real issues— the math was not up to par, it was hallucinating— but it was dramatically better than anything he had ever seen. And we started putting, essentially, all of Khan Academy on a nondisclosure agreement.

And we started, we started playing with it and seeing what was possible. And we started really, there's ways through the math, there's ways to reduce hallucinations, especially when you anchor it, all the content that we had been making for 15 years plus. And we see how this can take us even further in that direction of scaling what I was able to do, or scaling what a teacher assistant could do for a teacher, not only giving them the information, recommendations, but also help with lesson plan writing, grading, progress report writing.

So we started working really vigorously on that. And you can imagine the debates that were going on in our team. This was before ChatGPT came out or anything. We're like, well, this could be used as a cheating tool, there are real security and privacy issues here— what if a student asks how to build a bomb, what if they want to self-harm, what if it's just inappropriate? So we said, look, we don't want to shy away from this. Let's just turn to all of those fears into features and say, hey, what if we have a moderate AI that will active notif notify parents and teachers, something's going shady. What if we make all the— especially under 18 conversations— transparent to the parents, to the teachers, the educators. We’ll promise to make sure we'll steer it in a way that it will not will be a really world class Socratic tutor. Ideally, we're gonna get the math errors now. So that’s what we're working on. 

In November 2022, ChatGPT comes out. I always have to explain this to people? No, they hear anything at all with GPT kind of fires the same neurons in people’s head. ChatGPT when it came out, they were on GPT 3.5. And immediately people started seeing it as a cheating tool, they saw it was cool, but it had some issues. School districts started banning it— for good reason in my mind— and then I Slack Greg Brockman. And I was like, what, what's going on here?

You have us under an NDA, a non-disclosure agreement. But you just launched something and he said, “No, we didn't launch anything, we just put a chat interface in front of GPT 3.5, which had always been around and the whole world kind of exploded.”

I was super bummed. I thought that the education system was going to throw out the baby with the bathwater, and just not want to deal with that. But it was a blessing because by March of 2023, when we did launch with GPT4, which was a much better model. And then educators were essentially saying, we want the power of this technology, we just wish someone would do it in a way that is safe, that has guardrails, with someone that we know and trust.

Ideally, something that's not just a standalone wrapper around an AI model, but maybe it's integrated with curricula, or interventions that have actually been proven out. That's not pure AI-generated, that actually already has a workflow with teachers and students. And so when we launched Khanmigo, we got, honestly, a shockingly positive reception.

Ben Kornell:

We actually interviewed Kristen, who’s part of your team, on the pod and she was talking about when you were working on it, it was almost like you could see the future, but you couldn't really tell anybody what the future would be like. And there was a way in which you were kind of guardians of this technology for the good of people.

One of the things you've been thrust into is a leading role as an icon of Edtech and of AI in education, and in many ways, your are synonymous with AI Education. That's a huge barrier area for our field and a lot of responsibility. How do you think about that? And how does that weigh on you? I will just say, side note, I have just been so impressed with the courage of your team that put things out there and not have it always 100% right. It’s too important. We need to do this. How do you think about playing that role for our space?

Sal Khan:

First of all, appreciate that y'all perceive us that way. And hopefully, it's not by accident. I mean, just as you talked to Kristen. Our team- when people join Khan Academy, its not to become gazillionaires, they join Khan Academy because they're passionate about education. And the debates inside of our team were the debates that, frankly, should be happening in broader society around safety and privacy. And okay, well, “what if this AI can still hallucinate, is that okay?” and so we started creating a framework,  we have to believe that the benefits of whatever we launch far outweigh the risks.

But at the same time, that technology is evolving so fast, that we can't be afraid to at least put things out there, as long as we're confident that the benefits far outweigh the risks, and go start working with real teachers, and real kids in real districts, and we learn from it. And so we built up our user research team— we always had a very strong user research team, but now they're like embedding themselves with teachers, school leaders and students.

And I tell the team this all the time: It's very tempting in any organization-  for profit or nonprofit- to say yeah, we all have egos, you get a little bit competitive, to say “we want to win.” And I always remember, as a nonprofit, we're not there for ego. We're there because in theory, if we didn't exist, the world might not turn out the way that is consistent with our values.

And I've always felt education and probably health care are the two parts of the economy where market forces have not always led to the best outcomes. So we have a public school system. We just think everyone should get an education. That's why, hopefully, if someone’s bleeding on the doorsteps of the ER, they're not checking their insurance, they’ll treat them first.

An with AI, as  I've been telling you, and I hope I'm not just rationalizing, that if we don't show a way to use this stuff well, there are about 500 people right behind us— most of them actually with good intentions, I actually think— but they don't always have the same setup to act with good intent.

Many of them have venture capitalists breathing down their throat. Many of my friends are venture capitalists, but I know what happens after a year, when it's time to turn the profit. And maybe we want to monetize the data a little bit, maybe we want to turn the product this way, “hey, let's use the cheaper model, it's a lot cheaper to use GPT3.5 than to use GPT4 and well, you know, it might hallucinate five times more, but at least we'll make the sale, right?” Like, that's the type of thinking.

And you can lead to just a deeply fragmented market, which is not going to be good for teachers.

Like even before AI, people have complained that there's a lot of interesting tools— Khan Academy is one of them— but to log into 10 different places and then have to write a report about, “okay, I'm using that for reading comprehension, I'm using that for writing and using that for math.”

And so I tell our team, we have a responsibility to show, not only can it be done right, but you can innovate it in an ethical way, and that you can hopefully make it a really compelling experience for teachers, students, and parents.

And also show that it's efficacious! We have 50+ efficacy studies at Khan Academy, so it's enhancing those things. Then we are going to be the first: we already have some internal studies: it's definitely not doing harm, it's driving engagement. We're already setting up our data in place so that by this time next year we will have data about efficacy. I don't think anyone else actually cares about that yet. So you know, knock on wood. The AI will have a good landing.

Alex Sarlin:

Khanmigo is rapidly becoming synonymous with AI tutoring- you already have 95,000 students and teachers are piloting Khanmigo in U.S. classrooms and in your pilots, you over index on school districts with higher percentages of students from under-resourced communities. Tell us about what you’ve been seeing in your Khanmigo pilots, as well as your new academic essay tool.

Sal Khan:

Khan Academy is oftentimes most associated with math, and maybe math and science. But our goal, Mission free world class education for anyone, anywhere has always been all of the core academic subjects. And I've always said “I'm going to do this until the day I die. And we'll eventually get there.” And we already have a lot of humanities, we have high school level and AP level history and civics and we were going to build that and- we are building for the launch this back to school - reading comprehension regardless. But if you asked me two years ago, when is Khan Academy going to have kind of top of Bloom's pedagogical pyramid writing, things like that? A long time from now, maybe forever. But now given the tools, it's not forever, it's actually now. And he's mentioned, we've already launched a tool where, because so much of the narrative around AI is cheating, especially cheating around essay writing, right? So we said, “Okay, let's show that there's another way to do this.”

So we already launched the tool to write your essay, they had the AI ism feedback the same way, if you were to write on, say, a Google Doc, and you're collaborating with a thoughtful editor, or writing coach, and then what we're doing well - what we’re launching for back to school is teachers can create a writing assignment rubric with the AI, assign it through Khanmigo, students work on it with Khanmigo - they get access to the coach, “what do you think your thesis should be?”, “let’s find some data to back it up”, “let’s just outline it.” It gives feedback, sends it to the teacher, could give a preliminary grade, save the teacher some time, but the teacher we think should always be the prime auditor.

But even more important, it's been supporting the student the whole time. So it can tell the teacher the entire process. So it can tell the teacher not just the output. But by the way, Ben had trouble coming up with his thesis, and Alex did too, and it's a pretty common theme and you might want to do this mini lesson for your students. And by the way, Ben's essay, he worked on it for four hours, like here's the whole transcript of us working on it. And it's consistent with his other writing. It's a little bit better because we're making progress, but it's consistent. I'm confident it’s Ben's work. But Alex's writing, it’s not consistent with other work, it just showed up, I didn’t work on this essay- you might want to double click on what Alex has been up to. And so we think that's actually the way that you undermine the cheating issue. But it's not about the cheating issue: it's actually about better support for the students and better information for the teachers.

Alex Sarlin:

Exposing the process.

Sal Khan:

That's right.

Ben Kornell:

Many people don’t know that Khan Academy’s support for schools and educators, because of the Genesis story around direct support to students and the ability to personalize, there's almost been a misperception that it was really around student engagement tools, but now there’s a whole host of schools and also districts can even pay for sponsorships to get your team to do professional development, the deep work. Can you tell us what you see the role is, will be, of schools and educators future, as we've seen more and more of what AI can do and what Khan Academy can do?

Sal Khan:

Yeah, as you mentioned when we got off the ground back in 2008, 2009, 2010, we were just making stuff— for a while was just me making stuff— and just putting it out there and people using it. But even from the early days, it was both student tools and Teacher Tools. As early as 2005, I was making reports for myself as my cousin’s tutor to understand what they're up to, and those have evolved into our art teacher tools over the years.

But you fast forward to about 2016. And even at that point, we had about 50 efficacy studies that all showed the same narrative that if students were able to put 30 to 60 minutes a week on Khan Academy doing personalized practice in math, that's where all the studies were- the students, depending on the study, were accelerating someplace between 20% and 50% faster than they otherwise would have. That's great. But we realize that as a nonprofit to actually move the dial for districts, states, countries eventually. So we started going to superintendents and started saying “look, look at the efficacy studies. And by the way, you already have 80 teachers in your district using Khan Academy. How do we work with the other 80 to use it?”

And with almost every conversation went something like, “oh, well, I believe you, because my son/daughter/niece/nephew uses Khan Academy.” For younger superintendents, some of them, apparently, we helped through grad school for their stats and whatever else.

So there was that trust and that belief. And I think that's important, because so many Edtech tools, they can do a nice presentation, but nobody has ever used it, the decision makers have never actually used the tools themselves. And so we view that as a very strong badge of honor. Someone can say, “Oh, I used it” or “my coworker swears by it.”

Then these, the next sentence that Chief Academic Officers or superintendents say, is “but if you want us to use Khan Academy for all of our classrooms, you have to get us support, training, integration with our rostering systems, district level dashboards.” And that was scary for us because we never had boots on the ground in that way, so to speak. But we realized if we're serious, we needed to do it.

But yes, Khan Academy always had real cost. It had historically funded our what's now almost a $70 million budget primarily from philanthropy, but for us to send people to a district and run professional development sessions and all that, cost a lot more money. So we told districts, we penciled it out, we're like, we just want to charge for that. And so that was $10 per year that we started charging them. And so that built our Khan Academy  district offerings, and then you fast forward and there's a million students in those districts that are doing formal partnerships.

And what we've seen is, it just gets things to a much higher degree of fidelity, you know, we started this conversation saying, What surprised you been, what’s been underwhelming? What’s surprised me is how many people do use Khan Academy, but underwhelming has been that fidelity of implementation has not always been strong. So sometimes it's just someone writing our URL on a chalkboard and hoping for the best.

But when we do these district partnerships, and we do the training, the support, the district officials can also see dramatically better engagement, the type of engagement, you need to see to see that type of efficacy improvement. And so now with generative AI we're gonna go to a whole other level.

Right now, it's not even just districts but states. The one I can talk about openly is the state of Indiana,

Ben Kornell:

I’m from Indiana.

Sal Khan:

Oh, you’re from Indiana? Well, Dr. Katie Jenner is a really great superintendent there. Once again, it's not her telling everyone, “you have to use Khanmigo.” She's like, “okay, maybe I can help the districts who want to do it use Khanmigo.”

So 30,000 of the 95,000 (Khanmigo pilot students) are in Indiana already, and I suspect by next year, it's going to be several hundreds of thousands. There's about five or six other states who are thinking about now replicating it, but it's all about: you build a model where not only is the tool accessible by the districts, but there's the right support for the teachers so that they can use it.

One of the things I'm really excited about with AI that's different than even free AI Khan Academy. Free Khan Academy, it's like look, if you use it in this way you can improve your outcomes, and most teachers of course want to do that, but they have so much on their plate. “Oh, but this is one more thing to learned etc. “

With AI, yes, it is one more thing to learn, but so many of the features- it's not just about personalized practice for the students or supporting them in writing, supporting them with talking to simulations of historical characters— it can also help with lesson planning, it can help with progress reports.

We have a school district out here in California where they're estimating they’re saving their teachers at least five hours per week per teacher. And so this is hopefully a refreshing Edtech message for a lot of educators to say, “Oh, this will actually save my time, it'll actually give a little bit of energy back to me that I can use for myself or for my students.” So we think that's also going to grow adoption.

Alex Sarlin:

I wanted to ask about your experience during the pandemic. Schools shut down, things were going crazy, so many of the families and students who are suddenly forced to learn from home are suddenly using Khan Academy, because it’s so ubiquitous. I’d love to hear your personal perspective about what it was like being inside Khan Academy during that time.

Sal Khan:

It was fascinating. I remember in February of 2020, a teacher from South Korea wrote to us saying, “Oh, we're using Khan Academy to get us through our nationwide school closure.” And I remember telling it, “this is unbelievable, a whole nation shut down its schools because of this virus?! And then it was literally that week that California, and then by the end, the rest of the country and world had shut down.

And so you can imagine our traffic went from- on an average school day, we were supporting 30 million minutes per day, it went to 90 million pretty much by the end of that.

Now, what was interesting, no one really understood at that time, and we still are probably the only player, or one of a few in which people are using it on their own, and we're working with districts and we're standards aligned, etc, etc, and we have teachers tools, so it was easy to switch.

As we went further and further into the pandemic, unfortunately, what we saw happen is, even pre pandemic, people had an, “okay, we're doing something now, what can we do to improve it, to make it better? Maybe we use Khan Academy, maybe we do this, maybe we do that.”

When the pandemic hit, a lot of people started using Khan Academy more intensively, and a lot of that might have been family driven. But then, once what was then called “COVID schooling” kicked in, which no one liked, no one: students hated it, teachers hated it, families didn’t like it. They just tried to transplant and- I don't blame it on them, how do you switch virtual learning almost overnight? But it created screen fatigue, it created a sense of, “hey, this isn’t time to innovate. This is just like making sure the car is moving.

Ben Kornell:

Survive and move on.

Sal Khan:

No one's to fault for that, what you saw was, I would say, a regression from any type of desire towards innovation. From “Let's make this more personalized,” to “it's a win if the kids are even in something that kind of looks like school, maybe test scores will slide” (and unfortunately, they did slide), and then that happened at the same time that everyone has screen fatigue.

Back in the day, it was for older kids, maybe like 15-20 minutes a day, that was not a ton of screen time. and for younger kids, maybe that two or three times a week, but not even. But when half the kids were already learning on Zoom for four hours, that's a harder ask. I think the whole Edtech market saw a regression. But then post-pandemic, when people finally started to name learning loss, unfinished learning- that's what we've been preaching for 10 years. And you should be able to fill this stuff. I think that’s the reason AI has started another tsunami of interest to drive adoption.

Ben Kornell:

Well, unfortunately, we’re almost out of time, but before we go, our listeners are Edtech entrepreneurs, they are leaders, schools, investors. Do you have any advice for others who are trying to drive impact for learning and kids using AI, what are some words of wisdom you would want to share?

Sal Khan:

I'd say two things. This is a time when technology is moving so fast. On one level, be bold and try new things out, but put them in front of real people. Don't just sit in a vacuum and think that because the technology is cool that the application will be cool. Put the pedagogy first and say, Hey, we personalize, we want to do mastery learning, we want to support teachers. Okay, now that you know the problem, then think about the technology. Maybe it's generative AI, maybe it's a pulley. It doesn't have to be high- tech, right? It might just be a way of thinking about things. But I would say as you do it, we should be very thoughtful about not making bad first impressions and putting the right guardrails in place.

You know, and right now in the broader AI conversation, I've been invited to some of these AI summits, not just education, but just broader AI. So they're bringing me because right now education is kind of the one place where they see that this is kind of positive. Everywhere else, we're rightfully worried about fraud and deep fakes and election manipulation and all of that, but Edtech seems positive. So what I would tell everyone else in the Education Committee, Let's not mess it up.

So let's make sure we're using the right models, let’s make sure we're doing it ethically, we're doing the right things on the data privacy side, that we're making sure it's safe, that we're making it transparent for teachers and students, that we're thinking about the teachers and the parents as a very important part of this loop. And then I think things could turn out pretty well.

Ben Kornell:

Wonderful. Thanks so much, Sal Khan. Incredible words of wisdom. This has been an amazing time for Alex and I, we’ve been huge fans. Sal Khan, CEO, Khan Academy. Where's the best place for people to find you on the internet? (Laughter) Khanacademy.com, will that work?

Sal Khan:

Khan academy.org. I think Khan academy.com points back to Khan Academy.org, but Khan academy.org. And people could just search for Khanmigo if you're interested in the AI stuff.

Ben Kornell:

Thanks so much for joining us.

Sal Khan:

Thank you.