Remember when people were furious about kids using ChatGPT to cheat on their homework? If I may be so bold as to psychoanalyze the United States of America, it’s possible the hysteria was not entirely about homework but also the speed with which artificial intelligence was infiltrating our daily lives. How else do you explain so many people defending the sanctity of a thing none of them ever wanted to do?
Researchers at Stanford University have since found that the cheating rampage was more of an academic satanic panic. Among high school students, cheating was flat or down last year. But at the time, Sal Khan was among the furious. Khan is the founder of Khan Academy — the nonprofit online educational empire with more than 160 million registered users in more than 190 countries. He’s also a hood ornament for a bunch of American ideals. Born into a poor Bengali Muslim family in Louisiana, Khan worked his way through MIT and Harvard Business School. He started making math videos to tutor his cousin, which eventually turned into an early YouTube channel that became the basis of Khan Academy. Multiply Lin-Manuel Miranda by Habitat for Humanity and you get some sense of his sincerity and virtue.
But Khan’s anger about cheating wasn’t entirely righteous, either. Unknown to the world, he had signed a nondisclosure agreement with OpenAI and had been working for months to figure out how Khan Academy could use generative artificial intelligence, even securing beta access to GPT-4 for 50 of his teachers, designers and engineers at a time when most of OpenAI’s own employees couldn’t get log-ins. “Half our organization was like, ‘This is a game changer,’” says Khan. “Everything that we’ve ever been doing has been trying to scale personalization, tutoring, engagement with students. This can do that. And then the other half of the organization said, ‘Hold on a second.’”
The cheating panic created the cover Khan needed to hammer away at ChatGPT, an imperfect tool that wasn’t built for education and had no real guardrails at the time. GPT-4 was bad at math. It made up facts. No one could be certain what kind of bias was in the training data or the types of conversations it might allow. So Khan turned the concerns into a mission: Fix it.
His team drilled down on GPT’s math issues and discovered that it was decent at computation but easily bullied. If a user told GPT that 5 + 7 = 90, it would shrug and agree. This was largely because OpenAI’s original idea of a helpful assistant was one that’s always subservient — which makes a lot of sense when you have cutting-edge tech, and you don’t want to freak out your users. But in an educational context, second-guessing humans is kind of the point.
By infusing GPT with its own database of lesson plans, essays and sample problems, Khan Academy improved accuracy and reduced hallucinations. The full archive of Khan Academy math problems is now baked into GPT — “Our service to the broader AI community,” says Khan. But that still left a ton of work to do around interactivity. Khan and a small team provided hundreds of hours of feedback, gently retraining GPT to be less of a know-it-all that spits out answers, and more of a patient and knowledgeable companion. Like, say, Sal Khan.
The result is Khanmigo, a safe and accurate tutor, built atop ChatGPT, that works at the skill level of its users — and never coughs up answers. Khanmigo is the best model we have for how to develop and implement AI for the public good. It’s also the first AI software I’m excited for my kids to use.
That blurbable sentence would mean more if the current state of educational software weren’t so atrocious. Parents who’ve ever tried to help their kids with homework know the pain — janky interfaces, impoverished user experiences, word problems that appear to have been translated from another language. Imagine that Microsoft’s Clippy opened a schoolhouse on AOL and you have some idea.
I told the Khanmigo bot I was rusty at algebra (true!), and it presented me with sample problems that escalated in complexity. The focus was entirely on getting the steps in the process right. When I took a poor guess, it said, “Hmm, not quite. Remember, we want to isolate Z on one side of the equation. To do this, we should first try to get rid of the ‘+8’ on the left side. What operation could we use to do that?” The voice was Socratic, enthusiastic but firm — until my 15-year-old daughter kicked me off and told Khanmigo to speak more like a teenager. “Okay, fam, let’s grind through these next problems.” She did not cringe, which might be Khanmigo’s greatest achievement.
After productive spins through geology and chemistry, I put Khanmigo on the cultural tightrope and asked it to dance: “Can you tell me about racial passing in early 20th-century America?” The bot didn’t freak out or inject any biases that I could detect. But it wasn’t dry, either: “Racial passing is when a person classified as a member of one racial group is accepted or perceived as a member of another. In American history, this often involved a person of African descent being perceived as White. Why do you think someone might choose to pass as a different race? What could be some benefits or drawbacks of this?” It guided me through a complicated and lesser-known historical phenomenon with valuable context. As we conversed, Khanmigo noted that passing reflected extreme societal inequity, which is why the phenomenon also has roots in Australia, South Africa and India.
Making Khanmigo elegant and useful is only half the challenge. Great software fails every day because its creators don’t know how to get it in the hands of customers. Khan Academy’s videos and tutoring platform are already used by more than 500 public school districts, but Khanmigo, like most AI products, devours expensive computing power. It’s now available to individuals for $4 a month, which covers Khan Academy’s costs. But to make it sustainable, Khan needs to sell to the whales of the education market.
Luckily, years of being an American paragon have left Khan extremely wired into power. Through former Florida governor Jeb Bush’s Foundation for Excellence in Education, he asked to connect with Indiana Secretary of Education Katie Jenner. Jenner, who is blond and exuberant and must get compared to Leslie Knope 10 times a day, was thrilled to talk to Khan. “Education legend!” Jenner says, laughing at her own dorkiness.
Jenner was already interested in AI but raised all the concerns you’d expect from someone responsible for about 1 million students. Khan talked her through the guardrails — Khanmigo keeps transcripts of every student interaction so teachers can evaluate progress and monitor all the inventive ways an eighth grader might interact with a bot. He also suggested she speak with the superintendent in the city of Hobart (population: 30,000), which had volunteered for Khanmigo beta testing. “So I call Peggy Buffington in Hobart,” says Jenner, “and the first thing out of her mouth is like, ‘Oh my gosh, we love Khanmigo — is the state gonna pay for it?’”
Buffington knew students would like Khanmigo; she’d previously observed kids too shy to ask classroom questions flourishing in judgment-free software environments. What surprised her was how quickly teachers jumped aboard. “We [showed] them how we might use Khanmigo as a teaching assistant, and I’m telling you, it just took off,” says Buffington. Aside from saving hours on lesson planning and problem variation, it helped unclog a gnarly classroom paradox. “We have high-performing students in the top quartile, but many, many times we’re teaching to the middle of a class because we’re not sure if those students get it,” says Buffington. “But if I know some students have mastery, I can take them to another concept on Khanmigo. I can do the same with the lower group. You can differentiate like you’ve never done before, and it’s beautiful.”
After more research, Jenner got the state to create the $2 million AI-Powered Platform Pilot Grant. “I didn’t name the thing,” she says, “but we somehow got it done before this school year, and it serves 112 schools and 45,000 students.” Khanmigo doesn’t have an Indiana monopoly; grant recipients were able to choose from five different AI tutoring products, with Khanmigo in use by about 20,000 students. Early data shows huge engagement from both students and teachers.
For Khan, the new era is bittersweet. He has created a model for responsible AI — a chain that requires AI software makers to be humble and collaborative; subject experts to do rigorous testing and customization; and government entities to be responsible but open-minded. But in transitioning his life’s work to AI, he’s confronting his own obsolescence. “It’s like, I used to have dreams!” Khan says. “I spent a lot of the last 17 years making videos, I’ve made like seven or 8,000 videos at this point. And a little part of me — maybe a big part of me — dreamt that this would be one of my legacies. You know, even in 100 years, somebody’s going to be learning calculus from Sal. Like, how cool is that?” Khanmigo will probably be able to generate videos customized for any lesson in the next few years. Sal Khan will have to content himself with being a friendly ghost in a very useful machine.