Sal Khan's $15 Tutor Hit 1.4 Million Users. Making Kids Actually Use It Is the Product

Jacob Effron opens with the question every AI show asks: what does a classroom look like in twenty years. Five years ago, Salman Khan says, that would have been boring. Now it feels too far out, because the ground is moving too fast. The picture he actually wants, even in five years, is not a sci-fi rebuild. Great classrooms already look like it. Students are not sitting through a lecture. They are solving, grouping, presenting. The teacher is on her feet, getting them to do interesting things, reading the room almost by intuition.

Those teachers are scarce because the superpowers sit on unpaid overtime. Great lesson plans, hours of grading, progress reports: they do the creative work *anyway*. Khan's bet is that AI can hand ordinary teachers some of that time back, plus better reads on where each student stands and how to run the room. In twenty years he does not think the laptop is the point. Generative AI can be ambient, watching the class instead of demanding a screen, so the old binary—technology is staring, non-technology is running around—stops being a trade-off.

Further out, virtual and augmented reality go mainstream, and with generative AI you get the field trip that sounds like fiction: drop into a simulation, walk ancient Rome, try to stop Julius Caesar's assassination or hurry it along. He calls it a Magic School Bus ride. The thread Effron pulls is the one Khan wants: not replacing the best teachers, but giving more of them the same superpowers.

Khan Academy's true north, which he says he has articulated more clearly since generative AI arrived, has not changed since the hobby years: approximate what a good tutor does. Family members first, then software, then videos, then a platform serving more than 150 million learners in 190 countries. Teachers with thirty kids still need a way to personalize. In the early days he kept citing Neal Stephenson's *The Diamond Age* and its Young Lady's Illustrated Primer. That is what they were building, incrementally, as the technology caught up.

The catch-up moment was specific. OpenAI gave the team GPT-4 months before ChatGPT existed. It had issues, but it was close enough to approximate tutoring and a teaching assistant, plus uses they had not conceptualized. They launched Khanmigo with guardrails: teachers can see what students are doing, it is built not to cheat, safety and privacy are in the spec, and the pedagogy is Socratic rather than answer-spitting. At launch he thought maybe 100,000 people would be on a pilot by 2025. It is now about 1.3 to 1.4 million teachers and students.

The mission is still free world-class education for anyone, anywhere. Compute, support, and training forced a price: about $15 a year to districts. A year and a half in, that many seats are paid, and interest is still strong. The first weeks after the GPT-4 demo, which Effron asks about later, were almost daily calls with OpenAI lawyers to put five or ten more people under NDA. Khan would pull colleagues into a small office and run the demo; people thought they were being punked. About half the company said drop everything. The other half listed errors, hallucinations, bias, freak-outs, cheating. His line, which he still uses: both sides are right, and the job is to take those risks and turn them into features.

The next phase, he says, is making the AI proactive. Even if Khan walked into a math class, announced he was a great tutor, and sat in the back, maybe 10 or 15 percent of kids would come ask. They are seeing the same pattern with the AI. The version they will pilot at back-to-school is Khan Academy Classroom: every visit greets you by name, reminds you what your teacher assigned, offers help. For the teacher it is a concierge, front and center, not a box you have to think to open. Effron calls it the blank-screen problem that haunts every AI product. Dynamic suggestion bubbles and training help. The core, Khan says, is the AI going first.

He throws cold water on the rest of the category. About 500 people claim to be building some version of an AI tutor. He does not think the models are ready to drive learning by themselves for most people. A curious adult who spends half an hour a night with ChatGPT, prompting well, can learn a ton. That is not most students. The systems are also not yet great at writing high-quality, low-error practice items, even if they improve every day. In Newark, where they did a large rollout, the striking efficacy numbers come from students doing traditional Khan Academy practice. The AI is support. It helps with engagement.

Edtech, he says, is always hunting for a more efficacious intervention. If the method is reasonably healthy, engagement is what actually works. The hard part has always been getting people to engage. So they look at the whole pipeline: activate the teacher faster, let district administrators politely hold classrooms accountable for using the tool, then let leaders and teachers hold students accountable. Human systems, with AI helping humans hold other humans to account, are how you get the hours. Later he is blunter about the mechanic. If the AI notices Jacob is stuck on the distributive property, it tells Mr. Khan, click here, assign a Khanmigo session, due tomorrow night. When teachers assign AI tutoring and then check that it happened, kids finally engage.

Effron has heard that students explaining their reasoning can pull the model into a useful loop. Khan says the truly magical cases are still not mainstream, but they exist if someone is motivated. In a TED talk two years ago he told the story of a girl at Khan World School, in India at the time, reading *The Great Gatsby* and having a long conversation with an AI Jay Gatsby. Their literary simulations do not just answer. They drive: what did you think, is there a moment in your life like that. Hearing her describe it, he thought: this is the thing we want more of. His oldest son, just sixteen and doing math at or past Khan's own level, uses the tools constantly to explore ideas. Asked if the AI is always right, the boy said no, but neither are you.

Khan Academy is spending real resources, more than he thinks most others are, on measuring accuracy rather than just improving it. Anchored on their own content, Khanmigo is at about a 2% error rate, split evenly: 1% straight math, 1% evaluation. The student writes 0.33 for one-third, the AI says great job, when it should have asked whether the threes keep going or whether a fraction would be better. Zero would be nicer, but when Khan tutors his own kids he is already wrong, or at least confused, one time in ten or fifteen. By that standard the machine is already better than a lot of human tutors. The hardest remaining errors are not fifth-powers. They are evaluation errors with students who hammer the bot, switch context, and fish for the answer.

On the teacher side the unsexy win is habit. Plan with the AI, punch up the lesson, deliver, read the data, plan again. A partnership with Blooket, the in-class question game that Khan says is as popular as Kahoot or more so, lets Khanmigo generate the set in about two minutes instead of the half hour or hour it took to write fifteen or twenty items. Sexier: open class by interviewing an AI Harriet Tubman or George Washington, hardest questions welcome. Writing Coach is their answer to cheating. The teacher builds and assigns through the AI; the student writes with it as an ethical coach; the teacher gets the process, not just the essay, and can ask the AI what happened. Paste from ChatGPT and it says it does not know where that came from. They believe that undercuts cheating in general, not only AI cheating.

Should school be with the tools or without them? Both, he says. If you are going to manage an AI at work you need to write well yourself, so pre-high-school and early high school should still do short, in-class writing with a teacher in the room. Later, use the tools for productivity. He has commencement addresses coming. In an airport he dumped life advice into his phone, had it transcribed, then turned into a first draft. It saved a mass of time and took him from something like 0.3 to 0.7 because the thoughts were already his. If he could not write and read that draft on stage, it would have been a horrible speech. The phrasing was sometimes beautiful and not quite him. People should learn that loop. Effron imagines the old two-week term paper shrinking to output by the end of the period.

Boards and CEOs are shoving AI into companies. Are schools under the same heat? Khan thinks schools may be among the first places you see mainstream AI for productivity and for the daily work of helping kids learn. Teachers already leaned in because planning and grading compress so cleanly. Compute is not free, but it is dramatically cheaper than what came before. After the pandemic, $86 billion in ESSER funds went out to help students remediate, on the order of more than $2,000 per American student. A lot of districts plowed it into live tutoring at $25 to $50 an hour. There were exceptions. For the most part, he says, there is not much to be shown for it. Instead of that hourly rate you are looking at $10 or $15 a year, with as much dosage as you want.

He is seeing more urgency in districts than inside Khan Academy, where he keeps asking when bookkeeping gets automated and why engineering is not 100% more productive like the company down the road. Some districts say the tools save teachers at least five hours a week, and they use that as a recruiting and retention pitch. Building it was not a thin prompting layer, which he thinks describes too many AI apps. Safety and moderation came first, then a climb down from over-conservative false positives. Math accuracy, especially those evaluation fights with answer-hunting students, took serious work. So did a UI that feels like more than a chatbot: Writing Coach with thesis brainstorming, an outline the model can see, a Google-docs-like draft with highlights on the dimensions that matter. The world, he says, does not want twenty AI apps. It wants one that chains and swaps prompts behind the scenes and keeps memory robust.

Asked what he still needs from the labs, he lists memory he can actually use and reset, advanced voice wired into the platform, and better generated questions. He has made about 7,000 videos; it breaks his heart a little that Khan Academy's written content may matter less over time, even as it trains the models. Video and image generation are fun without good pedagogical use cases yet. The demo he does want, maybe a year and a half out, is a tablet that watches a student show their work and talks them through it—the OpenAI demo from the previous year, which he notes was a fifth take full of errors until it worked. If that becomes robust and mainstream in a year and a half or two, it starts to feel like tutoring his cousins, or his kids. Duolingo can mint dopamine because it writes its own standards. Once Khan Academy went mainstream in schools, they had to align to standards that are not always pleasant. Learning still needs productive struggle. Quest-based shells—scavenger hunts, escape rooms—might smuggle hard problems into a game. He used a reasoning model to build most of his ten-year-old's birthday scavenger hunt. It did not work out of the box.

He no longer probes new models with an AP Biology exam. Gemini, Claude, Grok, and GPT-4.5 sit in the same plus-or-minus; prompt them well and you can get nearly identical behavior. Differences are subtle enough that he leans on Khan Academy's own evals: several hundred tough cases, including 0.33 versus one-third and the distributive property. At the start, of about 200 hard cases, the models failed roughly 70%—not 70% of live traffic, 70% of the nasty suite. Last he checked, failure on the very hardest was under 10%. They also machine-label interactions, then correct with humans. About six months before the taping they labeled some 2,000 sample conversations, which is where the 2–3% error numbers come from. They are now labeling not just errors but whether a conversation was productive, versus a kid saying "I don't know" on a loop. That part is still all human. A traditional classroom, he suspects, would show similar checked-out ratios.

For students who are already curious, Khan says, this is a dream: any question, any time. Efficacy work is starting to show a correlation with engagement and maybe with learning, but the models have to sit inside the learning experience, not just answer questions, before he expects real movement. Effron puts the San Francisco line to him: if the assigned-session loop is so compelling, why not dump the models on every child in an emerging-market school today. What are the techno-optimists missing. Some percentage of students can run with anything. Sugata Mitra's Hole in the Wall put tablets in rural India and claimed kids just started learning. Khan grants the curiosity and calls the coverage limited. Hand ChatGPT to village kids and a subset will use it well. If you do not know what you do not know, you would not even prompt it to walk you through the Indian national standards in algebra. Structured content is still the point of Khan Academy.

He had just met a freshman at MIT from Afghanistan. She could not go to school. Khan Academy was her education; Schoolhouse.world, a sister nonprofit he helped start, let her prove what she knew; MIT accepted that and admitted her. People already run with what exists. In another year or two he thinks you could get pretty far without a local algebra teacher if you had Starlink, a $30-to-$50 tablet shared by five or six kids over a day, and a path. Within about two years he hopes Khan Academy can issue high school credits and diplomas—transcript automation needs funding—and later college credits. College diplomas are further out. An internationally recognized high school diploma changes the reason a kid in India fights for device time.

The broader AI-education market, he says, is noisy: startups wrapping ChatGPT in a prompt layer, a little traction, almost no switching costs. Investors, he tells a room of investors, often cannot take the longer view. A startup that misses product-market fit in nine months disappears. Khan Academy is a nonprofit. If the market fully solved the problem it would not need to exist. Its value in this moment is permission: funders will give a five-year grant to think through assessment and writing while the ground moves, and schools trust that the org is not trying to undermine teachers, that it cares about pedagogy, that it has efficacy studies. That trust is an asset in a jittery AI market. Corners he wishes someone would nail: interviewing, which is expensive non-standard assessment. He used to joke that losing a Google final round to a coin flip should have come with a certificate so Microsoft did not have to start from zero. Khan Academy posted a PM job, hundreds of applicants, maybe twenty or twenty-five phone screens, and he is sure there is talent in the 280 they never call. Corporate training—cybersecurity, harassment—would be better as an AI simulation that stops and asks what you do next.

The skill he wants schools to train is what economists call entrepreneurship: existing resources, new permutations, value that was not there. It used to mean starting a company. Over the next twenty years it is table stakes in almost any career—hear about a tool, try it, stack it with the work in front of you, get to 80%, then use your own craft to finish. Organizations without that gene will not innovate, and their cost structures will be worse. In the quickfire he says the mind-change of the last year is obvious in hindsight: he would not have claimed AI alone solves everything even two years ago, and he has shifted further toward using it so teachers can hold students accountable. He already uses it to ask the "dumb" questions before recording a video, to make images, to dump a speech and take a first pass. He wants to try vibe coding; his sixteen-year-old, who builds games, refuses like an old man, calls it cheating, and insists on learning it properly. Khan's reply is that you have to do a bit of both. The surprise in shipping Khanmigo is how much work a magical model still takes at classroom scale, plus a lingering fear that ChatGPT leapfrogs the specialized app. The better surprise: the team now sees itself as AI-first, and the education world moved faster than he expected.