Sal Khan Thinks Every Child Can Finally Have a Tutor

Sarah Guo introduces Sal Khan as the person who turned YouTube chalk talks into Khan Academy, now reaching 150 million learners, and as someone still energized about personalizing that work with AI. Khan’s own account starts much smaller. His background was tech, then business school, then a job as an analyst at a small hedge fund in Boston. In 2004, a year out of school and newly married, family visited after the wedding. Conversation turned to his 12-year-old cousin Nadia, who was being placed into a slower math track. Her parents had not gone to school in the United States and, Khan believed, did not see the implication: on that track she would never reach calculus in high school.

He offered to tutor her remotely from Boston while she went home to New Orleans. The actual snag was unit conversion. She caught up, then got a little ahead. He called the school and asked that Nadia be allowed to retake the placement exam. They asked who he was. He said he was her cousin. They let her. The same student who had started as, in his phrase, a remedial case was moved onto an advanced track. He was hooked. He started tutoring her younger brothers. Word spread through the family that free tutoring was available. Before long he was working with 10 to 15 cousins on top of the day job.

The pattern was consistent. They struggled because they had gaps. An algebra equation went badly because they were not fluent in dividing decimals, negative numbers, or exponents. For fun, because that part of his brain was idle at the fund, he wrote software that generated problems, gave immediate feedback, offered hints, and sat on a small database so he could see what each cousin got wrong. The domain Khan Academy was available, so he registered it. It was a family project. In 2006, after a chain of accidents he half-jokingly attributes to benevolent aliens preparing humanity for first contact, his two-person fund’s boss had a wife who took a law professorship at Stanford instead of UVA. Khan and his physician wife landed in Silicon Valley. At a dinner party a friend asked how he was scaling the lessons. He was not. The friend said there was this YouTube thing. Khan thought that was a horrible idea. YouTube was dogs on skateboards and cats playing piano.

He found screen capture, recorded the explanations he kept repeating, and put them up. The cousins famously told him they liked him better on YouTube than in person. What they meant, he thinks, is that they liked an on-demand version with no judgment and no shame if they had to review, available all the time, even as they still wanted him in their lives. Then strangers arrived. He had to shut down registrations on the web app because it was crashing his $30-a-month Java servlet host; his cousins could not get in. The first school to use the homebrew software, in 2007, was Sidwell Friends in Washington, the elite campus of senators’ and presidents’ children, in a fifth-grade classroom. By 2008 about 50,000 people a month were watching the videos. He still liked the hedge-fund job and the pay. He set Khan Academy up as a nonprofit with the mission of a free, world-class education for anyone, anywhere, thinking he might hire a few people on the side. A year later it had taken over his life. Real wealth, he told himself and his wife, is being able to do what feels like your purpose, as long as you can put food on the table, go to a restaurant now and then, and hopefully send your kids to college. They were trying to save a Silicon Valley down payment. He took the plunge anyway.

Guo asks whether, at some point, he actually studied education, the two-sigma problem, pedagogy. In the early stages, Khan says, it was intuition. He puts himself, and his hosts, in the category of kids who did well in school. You can tell yourself you are gifted, or you can notice that peers who are flunking the same class can beat you at chess and solve a puzzle faster. He used to tutor classmates in high school and saw the same thing he later saw with his cousins. The advantage the successful students had was that they did not allow gaps to sit there. When they felt uneasy about fluency they gave themselves extra practice, and they were not just memorizing. They wanted to know why the long-division algorithm worked, why place value worked. Later math then got easier, not harder. Do the opposite, skip the concepts and carry fluency gaps, and the cognitive load eats you.

He has walked classrooms where seventh graders understand exponents and still grab a calculator for three times seven. He is not exaggerating: he saw it in the Bronx, slapped the student’s hand, and made the kid work it out, fingers and all. On some level, he says, it is malpractice that nobody sat with that child for an evening and said this has to be automatic, in milliseconds. The boy understood the concepts of exponents and algebra. He did not have the working-memory headroom to use them. Educators later told Khan he was talking about mastery learning: always having the opportunity and the incentive to fill gaps. They told him he was talking about differentiation. Benjamin Bloom had coined mastery learning and, in 1984, written the two-sigma problem: one-on-one tutoring and personalization can accelerate a student dramatically. Two standard deviations takes someone from the 50th percentile to the 96th, an average student to an exceptional one. Bloom called it a problem because no real classroom can give everyone a tutor. Khan’s intuition, he now knows, sat on a large body of literature.

A textbook strategist would cringe at how wide the charter is now. The mission is still a free, world-class education for anyone, anywhere, which sounded almost delusional when he was one guy in a closet. Khan Academy Kids is used by nearly two million children, mostly in the United States, averaging about 90 minutes a month, and he describes it as turning into the Sesame Street of this generation, covering math, reading, writing, and social-emotional development for ages three to seven. The main site, “big Khan,” runs from the basics of math through calculus, statistics, and multivariable, plus science from late elementary through early college, humanities, art history, financial literacy, computer science, and economics, with most of the resources sitting in an exercise platform of deep item banks, immediate feedback, and teacher assignment tools. In 2014, unwilling to send his oldest into a local kindergarten that was not doing mastery learning after writing *The One World Schoolhouse*, he started Khan Lab School on the first floor of the Khan Academy offices. It is now K-12, with three graduating classes. With Arizona State he launched Khan World School, sixth through 12th. There is a pilot with Howard University in which students at Title I high schools who reach mastery on Khan Academy earn college algebra credit. And there is Khanmigo: in their minds, a tutor for every student and a teaching assistant for every teacher.

Khanmigo can sit with a student on ordinary Khan Academy work, answer questions, make the material relevant, motivate them, act as a guidance counselor or academic coach, debate them to sharpen an argument, and let them talk to simulations of historical or fictional characters. Teachers can use it to create lesson plans, grade, write rubrics, and refresh their own knowledge. Asked how important AI is relative to YouTube and online coursework, Khan splits the horizon. In the very short term it is a meaningful add-on. Three to five years out, it is a game changer. The gold standard was always a personal tutor. Alexander the Great had Aristotle. Mass public education was a utopian idea that, to be affordable, batched children into groups of 30, moved them on a timer, lectured, tested, and sent the ones who missed 20 or 40 percent of the material on to harder material anyway. Gaps accumulated. Kids fell off and got tracked onto different assembly lines, which was almost acceptable in an industrial economy that did not need a large knowledge workforce. It is not acceptable now.

Bloom, in 1984, showed the two-sigma gain from one-on-one tutoring and conjectured that approximating a tutor with technology might buy you one standard deviation. Everything Khan Academy had been doing was scaling pieces of that tutor: a video as an on-demand micro-explanation, exercises with immediate feedback, dashboards so a teacher can pull three or four kids who need negative numbers while the rest keep going. Generative AI is the piece he did not think would arrive in his lifetime. Playing with GPT-4 about a year before this interview, he was stunned that it could not only pretend to be a tutor but had good tutoring moves, that it was truly Socratic. Math errors and hallucinations were real; they have been working feverishly to mitigate them. It is already in real schools. Anecdotally, more kids stay engaged and get less blocked. Teachers say it answers questions students were afraid to ask, or that the teacher would not have reached. The assistant features save a lot of time on lesson plans and rubrics. Teachers feel less alone.

The same day as the taping, Khan Academy announced a partnership with Instructure, the company behind Canvas, which reaches most of higher education and a large share of K-12. If the AI creates the rubric and then administers the assignment alongside the student rather than for them, it can report the process, not just the grade: they brainstormed thesis statements, the student got stuck backing one up, they got there, it took about four hours. In that situation you can be reasonably confident the kid did not paste from ChatGPT. In a year or two he expects a useful tool that raises teacher productivity. Further out they are prototyping memory. If a student says they love women’s soccer, the next session should not have to ask again; if they prefer casual language or a given reading level, remember that too, and show it to the user so they can change gymnastics for soccer. He knows it sounds wild, but in five or ten years he can imagine a student’s best college recommendation coming from Khanmigo, which has worked with the child for 12 years and happens to know they love gymnastics.

Guo asks whether he has read Neal Stephenson’s *The Diamond Age*. He used to give a copy to every Khan Academy employee. In the novel a piece of neo-nobility commissions an AI tablet, the Young Lady’s Illustrated Primer, for a granddaughter. It is bootlegged into the hands of 200,000 orphan girls living on barges, who essentially take over. That was always the long-run picture for the team, and he never thought they would fully build it. Khanmigo already does things he puts beyond the Primer: talk to Don Quixote in any language, get into debates. In the next three to five years, he thinks almost everything Stephenson imagined in a book from 1995 is going to be reality, which he did not think would happen in his lifetime. Text-to-speech that is hard to tell from a person is coming, useful for tutors and scary on other dimensions. In three to five years you will video-conference with the tutor and it will feel like a long journey rather than on-demand help.

What happens to the actual school once the tutor exists? At Khan Lab School and Khan World School they have been running mastery learning even before generative AI. He is almost embarrassed to say the results. Steve Levitt of *Freakonomics* looked at Khan World School himself because he did not believe it, then confirmed it: the kids are learning about three to four times faster, not three percent or 30 percent faster. At Khan Lab School they have consistently seen 1.5 to 2 grade levels of growth in a year, also before generative AI. Traditional U.S. schools average about 0.7 grade levels per year. In more affluent areas, where parents are college-educated and can buy extra support, you probably see at least a grade level, maybe closer to the lab school’s 1.5 to 2. In those fluent environments, he thinks AI will free teachers. Most spend about half their time on grading and lesson planning. Give that back and they have more energy for themselves and more energy to be with students.

Three to five years out he pictures core fluency feeling supported, then something closer to *The Magic School Bus*: visors, a jump into the circulatory system, a time machine to ancient Greece for a debate with Socrates. Middle schoolers with time for passions might make studio-quality techno or a Lord of the Rings-scale fantasy film. In mainstream classrooms the nearer win is core fluency, and not only in math. Writing is resource-intensive. Feedback for 40 kids is expensive and mind-numbing; students wait a week and often never iterate, so most people do not get much practice. That is already a writing problem stacked on the math problem. In Newark, New Jersey, which he is careful to call not a special-case district, the North Ward started using Khan Academy intensively in November, still pre-generative AI. By June, 70 percent of those 6,000 kids were using it at a level associated with profound efficacy gains. Test scores were still out, but he thought they were going to show, for the first time in recent U.S. history, a large acceleration of a large urban district in math.

He has learned not to be too aggressive with predictions, but he thinks it is safe to say that in five years almost all teachers will use generative AI for lesson planning, grading, progress reports, and rubrics, and that it might happen by late the following year. Once teachers are already in the tool, a generative AI tutor for students becomes mainstream in three to five years. ChatGPT created an emergency around cheating. It is not a nice-to-have. People need a solution, and he thinks the solution is also generative AI, landing in the next year or two. Five years out a classroom might not look different on the surface. Double-click and teachers have more student-facing time, fewer students are stuck, and there is more room for differentiation. You will see it in engagement, motivation, and test scores. Teachers have a lot to gain. That is why he expects faster uptake than education usually manages.

Human connection is not optional in this picture. During the pandemic Khan Academy usage went up 3x, and isolation made the limit obvious. With Shishir Mehrotra, the founder of Coda, he started Schoolhouse.world, a sister nonprofit whose mission is to connect the world through learning. The bet was volunteerism: plenty of people would like to do what he did with Nadia. Tutors are near-peers, high school and college students, plus private-school teachers who want to give back and retired professors. Scale is much smaller, on the order of 10,000 people a month, but a large fraction of students who take any given SAT are already using it for live human tutoring. Khan World School’s daily anchor is a Socratic seminar, an hour or two of debate, not five hours on Zoom, which he thinks ruined pandemic schooling. Visit Khan Lab School and you see more collaboration than at almost any other campus, even in competitive Silicon Valley high school. Mastery learning, in his telling, removes the need to elbow a peer. Everyone here should learn it.

The GPT-4 relationship is another data point in the benevolent-alien theory. At the end of summer 2022 he got an email from Greg Brockman and Sam Altman: they were working on their newest model and wanted to talk. He was skeptical it would matter strategically, but he had followed GPT-2, GPT-3, and DALL-E, and he wanted a look. They had not even finished the first GPT-4 training run. They thought it would wake people up to generative AI, and they wanted to launch with positive social use cases. Education and health care were the obvious ones, and Khan Academy was the first call, because they wanted organizations people trust. The second reason was Bill Gates. After GPT-3, Gates had said he would be impressed when a model could pass the AP Biology exam, and Khan Academy has a lot of AP Biology items.

Two or three weeks later they demoed an osmosis multiple-choice question. Khan picked C. The model picked C. He asked it to explain why C was right; it did, quite well. He asked why the other choices were wrong; it did that too. He asked it to write five more questions like that one, and it did. That is when he got goosebumps. They gave him access. That weekend he could not sleep. He, the chief technology officer, and the chief learning officer were Slack-ing each other at four in the morning with things that felt like science fiction, most notably a Socratic tutor persona that would not just hand over the answer. They also found math errors, surfaced them, and learned there were mistakes in the training data. Then ChatGPT launched a few months later on GPT-3.5. He Slack-ed Brockman, thinking they were under NDA. He was bummed at first. In hindsight it helped. ChatGPT threw an imperfect, not-designed-for-school tool into the world, the education system took the cheating panic on the chin, and when they launched with GPT-4 on March 14, 2023, they could speak directly to concerns people already had.

Safety started from the cheating use case. Guardrails in the system prompt say do not give the answer, make the student do the bulk of the work, help by asking leading questions. They test and tweak those prompts constantly. Second, every interaction for a user under 18 is recorded and visible to parents and teachers. He knows older students might have constructive uses they would rather not share; they took the conservative stance anyway. Third, a second AI moderates the conversation. If it heads somewhere unhealthy, it shuts that thread down and actively notifies parents and teachers. They are also teaching digital literacy: here is what generative AI is good for, here is where you should be skeptical. None of the interactions are used to train the model, with OpenAI’s agreement. There might be future benefit in training on them. Because this is a sensitive frontier, they chose to keep it as kosher as possible.

Generative AI, Khan argues, does not dramatically change what was already going to happen to universities. The return on investment, especially with debt, is mixed. There is a student-debt crisis, and forgiving some of it does not fix the next generation. Universities are bloated and uneven at preparing students. The very elite ones function like a caste or a club; they will be fine whether or not they can justify the ROI, though legacy admissions will stay interesting. Community colleges at the other end are flexible, low-cost, and already thinking in industry certifications and mini-badges; they will be fine too. The trouble is the vast middle, especially campuses charging $60,000 or $70,000 a year with room and board, graduating students into debt that is hard to discharge in bankruptcy without the pay to match. Someone lent the money, the university took it, and now a trillion or two trillion dollars is ruining lives. If universities were on the hook for even 10 percent of that debt, he says their tune would change overnight. They would start telling students the career outlook. Not everyone becomes a museum curator, and the one who did may have had a grandmother who built the museum.

He wants a competency-based world: not how long you sat in a chair or went to frat parties, but what you know. Khan Academy and others are working on credentials employers will recognize, whether you earned them at 16 or at 36. That does not “disrupt college.” It gives college an alternative, which he thinks is healthy. The Howard University algebra-credit pilot is a small version of the same idea. Guo and Gil, like Khan, have preschool and kindergarten-aged children, so the skills question is not abstract. He offers two metaphors. The better one is not steam engines and horses. It is the camera. Portrait painting was a technical field whose prestige sat in accuracy. The camera captured reality better than anyone, and artists declared the end of art. Impressionism arrived in the same window. Art moved from capturing reality to conjuring a feeling, and it escaped the trap of painting nobility. Writing, code, images, and movies that machines can now produce, in his view, likewise liberate the creator to elevate, integrate, and manage the tools.

The second metaphor is the assistant. If every teacher gets a teaching assistant, every coder gets an army of coding assistants and every writer a writing assistant. People will have to move into managerial ranks quickly: coder to architect, writer to editor. To do that you have to know those skills really well, arguably better than previous generations. So it becomes more important, not less, to get not merely adequate but excellent at writing, communicating, and coding. Creativity plus these tools is, in his phrase, godlike. There will be a pile of terrible thousand-dollar movies, the way there were terrible hundred-million-dollar ones, and every now and then the Justin Bieber of filmmaking, a kid who made something extraordinary on a tiny budget. YouTube discovered talent that would not have been found otherwise, himself included, he adds, a little self-aggrandizing. For parents the advice is traditional. Get really good at reading, writing, and math. He has never subscribed to the idea that Google means you do not have to learn anything. The more knowledge you have rapid access to in your head, the more these tools accelerate you. The kid who needs a calculator for three times seven, or Google for when World War II was, is the one the world is going to pass by.