Sal Khan: "I see AI as an additional tool, but a very powerful one"

Interview by Anuliina Savolainen 
UNESCO

Your AI-powered tutoring tool is currently being tested in schools in the United States and online. Can you share with us some early impressions from students and teachers?

We launched Khanmigo on 15 March 2023, as part of the general launch of GPT-4. It was immediately  used by several thousand people, including students and teachers in Khan Lab School in Mountain View, California, in Khan World School (online) as well as in mainstream public schools in Newark in New Jersey and Hobart in Indiana. Now, as we go to this school year, tens of thousands of around 11,000 students and teachers are going to be using it in a formal classroom setting in the United States.

We first wanted to capture people's immediate reactions and to make sure we weren't causing any harm. Right out of the gate we heard very positive feedback from both teachers and students. We also have some preliminary data showing that it's definitely not doing harm.

The students really appreciated being able to ask Khanmigo questions right in the moment. We've all forgotten things and felt embarrassed to ask someone else – an on-demand video or AI tutor can feel less intimidating. 

The number of questions students were afraid to ask in class surprised teachers. They said it was useful to get a report back on these questions, so they could ensure such concepts were covered in more depth. They also appreciated using the tool for things like creating lesson plans and assignments.

Another thing that teachers and students have enjoyed is Khanmigo’s feature where students can talk to simulations of a historical or literary character.  Debating with AI allows kids to fine-tune their arguments in a safe environment before going back to class and entering discussion.

By the end of the school year we'll have some real data on what it is doing to students' learning outcomes. A recent study indicates that mainstream school students using Khan Academy for 18 hours over one year see their level growing 30 to 50 per cent faster compared to a typical student. We'll see what we can do with Khanmigo.

Amid fears of AI taking over education, how can you ensure that the environment you have introduced is safe and under meaningful human control?

The first immediate fear that people have with generative AI is that it can be used to cheat. This is one of the reasons why we use GPT-4 and not GPT-3.5, the technology behind ChatGPT. No matter how we tried to steer some of the earlier models, they would always just give the answer, and sometimes an incorrect one. Using GPT-4, we've been able to make Khanmigo act like a Socratic tutor.

Every interaction a student under the age of eighteen has with the tool is logged and accessible to parents and teachers. A second AI is monitoring the conversations and if they go into any ‘dangerous places’, the AI will not allow that conversation to go on and it will notify parents and teachers. We also make sure that any personally identifiable information like names or addresses is anonymized. We are not using student data to train the AI.

Generative AI is not always right, and it can sometimes make up facts. This is why most of the interactions that students have with Khanmigo are anchored on Khan Academy information, reducing the likelihood that it would go beyond that. On the maths side, we've done a lot of work to make it less overconfident. Khanmigo tries to find the answer on its own behind the scenes, and then compares it with that of the student. If the answers are different, it won't immediately say, ‘You're wrong’, but instead, something like, ‘Hmm, I got a different answer. Can you explain your reasoning?’. 

The last guardrail is making sure that the users, both students and teachers, are educated on what this technology is, what it can and cannot do, when you can rely on it and when you should double check its work.

Khan Academy promotes an  “open access free world-class education for anyone anywhere”. Khanmigo, however, is neither free nor yet accessible worldwide. What can be done to ensure that these tools also reach people living in remote areas and those with low income?

Today anyone in the US can sign up to Khanmigo. I think in the next few months we're going to be able to give access to it to anyone in the world who pays.

The generative AI costs depend on how much it’s used, but the average cost today is around 9-10 dollars per user per month. I predict that in the next year the cost will be at least half, maybe one quarter of what it is today. For rich countries like the United States, where the public school system spends between US$10,000 and US$40,000 per student per year, this is a very valuable thing – and hopefully the students will get it for free. But even if you think about the cost of education in other parts of the world, like India or Africa, I believe it becomes interesting. My hope is that in five years we can afford to give it for free or almost free to most of the world.

I see Khanmigo as an additional tool, but a very powerful one. In its current state, it works very well for curious students who want to make sure they have their conceptual gaps filled in. But hopefully it's also going to help disengaged students set and reach their goals. If you go to a school that has more resources, the teacher can sit one-on-one and do this. But in a traditional public school, where you don't get that type of extra attention, at least the AI can sit with you.

I do this with my own kids. I make sure that every day they get a little bit of supplemental learning. We want AI to get proactive like that. It's texting you, eventually it’s going to call you up on the phone and say, ‘Hey, I see you're not working today. What's going on? What can I do to get you working?’ As long as it's transparent, I think it can actually get students to engage.