Should Chatbots Tutor? Dissecting That Viral AI Demo With Sal Khan and His Son

Khan has become one of the most well-known boosters of using generative AI for tutoring, and he has a new book that makes an enthusiastic case for it. The book is called “Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing).”

But his book, and that demo, are also attracting some pushback from teaching experts who think AI may have lots of uses in education, but that tutoring should be reserved for humans who can motivate and understand the students they work with.

For this week’s EdSurge Podcast, we talked with Khan to hear more about his vision of AI tutors and the arguments from his recent book. And we also heard from Dan Meyer, vice president of user growth at Amplify, a curriculum and assessment company, who writes a newsletter about teaching mathematics where he has raised objections to the idea of using AI chatbots as tutors.

“The kind of math that we saw on there,” Meyer said, referring to the demo, “was an operational problem well summarized in a single diagram that results in a single number. And those have always been the kind of problems that computers have supported students fairly nimbly in solving.” The bigger question, he argues, will be how such chatbots will handle more conceptual problems. And, he asks how well such bots will work “for the average student who’s dealing with distraction and feeling socially isolated and not interested in talking to Scarlett-Johansson-esque voices as a tutor bot?”

Hear the full conversation on this week’s episode. Listen on Apple Podcasts, Overcast, Spotify, or wherever you listen to podcasts, or use the player on this page. Or read a partial transcript below, lightly edited for clarity.

EdSurge: When we last talked with you for the podcast, Khan Academy had just released your group’s chatbot tutor, Khanmigo. At the time you were rolling it out slowly because there were many questions about using AI chatbots in education. What was your biggest worry then, and how did the testing go?

Sal Khan: When we launched back in March of 2023, I think the biggest worry was how we would be received by the education community. This was only three or four months after ChatGPT had been released. And obviously the reception to ChatGPT was not a positive one, for good reason. It could be used to cheat. It had no guardrails on it. It was making math errors. It was hallucinating. And so here we are, an education nonprofit that hopefully a lot of folks trust to have high-quality work. And then people might say, ‘Hey, wow, Khan Academy is going with both feet into this AI thing.’

The good news is that the reception was actually more positive than we expected. So four or five months after the release of ChatGPT, most school systems, most educators were saying, ‘You know what, ChatGPT still is a little bit shady for education purposes, but the underlying technology of it is really potentially powerful for helping kids learn the things that we’ve always tried to teach them, and this type of technology is going to be part of their future. So we should think about how we can expose kids to it, but in a way that it doesn’t cheat, in a way that there’s guardrails, in the way that we can make sure that everything’s on the up-and-up on data security and privacy and that it’s pedagogically designed.’

And so when we were able to come with Khanmigo at around that time, the reception has been very positive.

I’ll say it’s also been a bit of a transition internally at Khan Academy because it is a new muscle that we’ve been building. … We’ve always worked on software that personalizes things, videos — I still make videos — and exercises, teacher tools, in a more traditional sense, and now we’re moving toward this artificial intelligence world. That is exciting, but it also has a lot of things to keep in consideration. I think it’s also been a bit of a transition for our team to feel good and confident and comfortable with where we’re going.

It sounds like Khan Academy will continue to make videos?

The rate of change of artificial intelligence is so fast that it feels like it’s irresponsible if we don’t have these conversations like, ‘How long will Khan Academy videos be relevant?’ A lot of folks probably saw the recent OpenAI demo of me and my son. Will a student find value in a Khan Academy video in that world, or as much value?

A lot of our resources historically have been creating these really high-quality exercises. We’ve created over 100,000 exercise items on Khan Academy, and that takes a lot of resources. Today, the AI is not good enough to create exercises that are high quality, aligned to standards and are error-free.

So AI is not replacing your job of making educational videos?

My vanity wants to say no, but I don’t know. I don’t know.

I do want to be clear. I think the safest job in all of this is that of the teacher. I make that very clear in my book, and I’m not just saying it because people want to hear it, but it’s that human element of it all being in the room helping guide students, keeping them on task, and you need to be physically there to really, truly keep them on task, to forge those human connections. …

But I think a lot of the other pieces that edtech has traditionally worked on or even other parts of the education system, maybe some of the more administrative tasks, I think it is important for everyone to be wondering how AI might change that.

You note in your book that back when you were an undergraduate at MIT, you originally wanted to be an AI researcher. Why were you drawn to that area?

I’ve always been fascinated by, ‘What could we learn potentially from technology?’ And I’ve always read a lot of science fiction books about maybe that could start pushing the frontiers of and even helping us understand what is intelligence and what is consciousness. But I’ve also been fascinated by the potential of human intelligence. And I’ve also always been fascinated by the intersection of the two.

And yes, when I was a freshman at MIT, I sought out for my freshman adviser and he ended up being my freshman adviser, Patrick Henry Winston, who was head of the artificial intelligence laboratory. I got in line to take a course with Marvin Minsky and got in. And so if you asked me in 1994 or 1995 what I wanted to do, I would say, ‘Yeah, I might want to be an AI researcher.’

Back then it sounds like you were discouraged by the level of technology at the time, but clearly we’re in a new phase of AI development. Do you think AI is now ready to serve as a viable tutor?

I think it can already do parts of it. I don’t think it’s able to do the full job, but I think that the technology is improving so fast that you definitely will never say never. And in fact, a lot of things that seem like science fiction are going to be reality in about two years.

[At Khan Academy] we’ve always been trying to use technology to approximate what a great tutor would do in terms of personalized learning and then also leverage technology to scale that to as many people as possible. And we’ve never viewed this as somehow a substitute for a teacher. In fact, we said, ‘Hey, this could be really valuable in a teaching setting.’ In fact, it’s most valuable in a teaching setting because a teacher’s in a class of 30, these kids are at all different levels. Every teacher knows that. How do you address their individual needs? Well, if you had support from a teaching assistant who’s also their tutor, that’s kind of what Khan Academy has always aspired to be.

I talked with a technologist who worked at IBM and had worked on IBM’s Watson many years ago and was asked to use it to build an AI tutor. But after years of work he concluded that it can’t be done, and that it’s not the best way to use AI in education. What would you say to that argument?

Actually when you talk to a lot of the AI researchers, and we’ve probably helped skew this conversation, the thing that they’re most excited about for the next generation models is the tutoring use case because people understand it’s a socially positive use case. Obviously there’s a bunch of negative use cases of AI — deepfakes, fraud, etc.

I think you’ve had many people work on this problem for decades using more basic forms of artificial intelligence. I encourage that researcher to watch that video of the GPT-4o tutoring demo with myself and my son.

Dan Meyer recently wrote that while these AI tutors might work for a small percentage of students, most need the kind of human relationship that just can’t be replicated with AI right now. Will a broad range of students want the kind of Khanmigo tutor you show in your demo?

I mean, I think most kids would rather chat or talk to their friends than go to school altogether, than sit through a lecture, than do their homework, etc. And this is why one of the many important things that a teacher does is make sure that students are focused and engaged on the thing that matters most.

There’s a broad group of students that, in the moment where they need to understand a concept, where this can be very useful for them. I agree that it’s a subset of students, let’s call it 10 or 15 percent of students who have maintained their curiosity and might automatically keep going to the AI. And for those students, this is a field day, this is a playground, this is awesome for them. I think there’s a broader set of students who are broadly disengaged from what they’re doing, and you need to figure out ways to engage them more. And this is one of the many reasons why we view involving teachers in this journey as so important. Letting them know what’s going on with the AI. We’re working on them being able to assign AI-based activities.

There’s a passage in your book where you describe Khanmigo having a session with a student and then reporting back to their teacher, and you write it might go say something like, “We worked on the paper for about four hours. Sal initially had trouble coming up with a thesis, but I was able to help him by asking some leading questions. The outlining went pretty smoothly. I just had to help him ensure that the conclusion really brought everything together. … based on the rubric for the assignment, I’d recommend Sal get a B plus on the assignment. Here is a detailed breakdown of how I rated this paper in the dimensions of the rubric.” In some ways, this doesn’t leave much left for the teacher to do. What would you say to teachers who worry AI could replace them?

I think every K-12 teacher will look at tenured professors at the local university with envy because those professors have a lot of support. They have these grad students who essentially do exactly what that example the AI was doing. So if you told every teacher in America, ‘Hey, we just found some money and we’re going to use it to hire some amazing teaching assistants that can help you write lesson plans, create rubrics, tutor your students, report back to you, what’s going on and do preliminary grading. You’re still the teacher, you’re in charge, but it’ll save you the teacher 10, 15 hours of your week. Do you want that?’ And I think the great majority of teachers will say, ‘Hallelujah. Yes, I definitely want that.’ I’m serious that I don’t think it in any way undermines the teacher. I think it elevates the teacher.

Back to that recent demo of the next-generation AI tutor. I’ve heard that your son already knew the material being asked and was sort of role-playing there.

Yeah, OpenAI said, ‘Hey, can you bring with you a student who can sign a media release who doesn’t work for one of our competitors?’ And I was like, I guess I’m going to bring my son. But yeah, my son, to his credit, he’s more low-ego than I am. I mean, he took calculus in seventh grade. He knows what a hypotenuse is. But it made a better demo for him to pretend that he did not know a hypotenuse is because it corrected him, etc.

But yeah, it is powerful to see it in action with a student where it can see what they’re drawing and what they’re saying, and it’s interacting verbally in a very natural way.

How long until the technology in that demo is actually fully functional in your tutoring chatbot?

I think we’re a year or a year and a half away from that. But even then to the earlier part of our conversation, even when it’s that awesome, I don’t know if every student in the world is just going to run to it.

We have a nonprofit called Schoolhouse.world, which gives free live tutoring over Zoom. But still, not every student who finds out about it runs to it. So the AIs are going to get better. There’s going to be other things like Schoolhouse World. But we’re still going to need engaged parents and teachers that can help motivate and drive kids to get the help that they need.