Salman Khan, founder of Khan Academy, discusses how AI is transforming education through Khanmigo, their AI tutoring assistant that has grown from an expected 100,000 users to 1.4 million students and teachers in roughly a year and a half. The conversation covers how AI can augment great teaching, why engagement is the core problem to solve, what it takes to build AI that works in real classrooms, and how education might look in 20 years.
The Vision for Future Classrooms
In 20 years, Khan believes classrooms will look more like what great classrooms already look like today: students actively problem-solving, working in groups, and presenting, while teachers circulate and facilitate rather than lecture.
AI’s role is to give all teachers the superpowers of the best ones: better lesson planning, real-time insights into student progress, and more time freed from grading and administrative work.
He expects AI to become ambient, observing classrooms and providing support without requiring students to stare at screens.
Virtual and augmented reality, powered by generative AI, will likely become mainstream, enabling immersive simulations like exploring ancient Rome or participating in historical events, similar to a “magic school bus” experience.
Khan Academy’s AI Initiatives
Khan Academy’s “true north” has always been personalization, inspired by the ideal of a great tutor, going back to its origins when Khan was tutoring family members.
He has long referenced Neal Stephenson’s The Diamond Age and its “Young Lady’s Illustrated Primer” as the aspirational model.
When GPT-4 became available (months before ChatGPT launched), Khan Academy saw it could approximate tutoring and teaching assistance, leading to the launch of Khanmigo.
Khanmigo was built with guardrails: it is Socratic, shows teachers what students are doing, prevents cheating, and prioritizes safety and privacy.
It launched as a paid product for districts at roughly $15 per student per year to cover compute costs, support, and training.
Adoption far exceeded expectations: from a projected 100,000 users by 2025 to 1.3–1.4 million within about 18 months, with continued strong demand.
Building Proactive AI
Khan observes that even a great human tutor sitting in the back of a room only gets 10–15% of students to engage, and AI tutors face the same “blank screen” problem: most students don’t know what to ask.
The next phase, being piloted as “Khan Academy classroom” for back-to-school, makes the AI proactive: it greets students by name, reminds them of assignments, and offers help before being asked.
From the teacher’s perspective, the AI acts more like a concierge, surfacing insights and suggestions rather than waiting to be queried.
Teacher and Student Experiences
Khan argues that for most students, AI alone is not yet ready to drive learning. A motivated person can learn a lot from ChatGPT, but most students are not that self-directed, and AI still struggles to generate high-quality, error-free questions consistently.
In practice, the strongest efficacy comes when AI supports traditional Khan Academy practice rather than replacing it. Engagement is the bottleneck: if students engage with something reasonably healthy, learning tends to follow.
On the student side, surprising use cases include students explaining their reasoning to the AI, which then iterates on it, creating a feedback loop that helps both the student learn and the AI self-correct.
A notable example: a student at Khan’s online charter school had an extended, Socratic conversation with an AI simulation of Jay Gatsby, exploring motivations and desires, which Khan described as exactly the kind of deep engagement they want to see.
Khan’s own 16-year-old son uses AI regularly for advanced math exploration and has pointed out that the AI is not always right, but neither is Khan himself when tutoring.
On accuracy, Khan reports that when Khanmigo is anchored to Khan Academy content, the error rate is roughly 2%: about 1% math errors and 1% evaluation errors (e.g., marking a correct answer as wrong or vice versa).
He considers this already better than many human tutors, based on his own experience tutoring his children.
Teachers are using AI in both practical and creative ways:
Practical: lesson plan generation, creating question sets for platforms like Blooket in minutes instead of hours, and using data from assignments to inform future planning.
Creative: opening class with AI simulations of historical figures like Harriet Tubman or George Washington and having students ask challenging questions.
Khan Academy’s “Writing Coach” tool addresses cheating concerns: students write with AI acting as an ethical writing coach, and teachers receive not just the final essay but the full process, including AI-flagged sections where the student may have copy-pasted from elsewhere.
AI Across Subjects and the Role of Tools in Learning
Khan believes students should learn to write and do math without AI, especially before high school, to build foundational skills.
At the same time, especially in high school and college, students should learn to use AI tools productively, because that is how the workforce will operate.
He gives a personal example: for upcoming commencement addresses, he recorded his thoughts, had AI transcribe and draft them, and then refined the output. The AI took him from a 3 to a 7 out of 10, but his own writing skill was essential to get from a 7 to a 10.
He imagines future classroom assignments where students are given a time-constrained task that requires using AI to produce output, reflecting real-world workflows.
District-Level Adoption and Policy
Khan sees schools as one of the first places where mainstream AI adoption for productivity and learning will happen, driven by the reality that so much of what teachers do (planning, grading, progress reports) can be streamlined.
Post-pandemic, roughly $86 billion in ESSER funds was spent on remediation, including expensive live tutoring at $25–50/hour, with limited results. AI tutoring at $10–15/year offers dramatically cheaper and more scalable dosage.
Districts report teachers saving at least 5 hours per week, and some are using AI tools as a recruitment and retention benefit.
Khan notes that as a leader, he is constantly pushing his own team to use AI for productivity, from bookkeeping to coding, and that schools are ahead of many other sectors in adoption.
What It Takes to Build AI for the Classroom
Khan emphasizes that Khanmigo is much more than a thin prompting layer on top of an OpenAI model. Significant additional work includes:
Safety and moderation systems, initially overly conservative with many false positives, now better calibrated.
Math accuracy improvements, especially handling evaluation errors with students who are trying to game the system or switch contexts rapidly.
A redesigned user interface to make the AI feel natural and integrated rather than like a chatbot.
Writing Coach, which includes brainstorming, outlining, and drafting tools with the AI providing contextual feedback on highlighted sections.
Prompt chaining and orchestration behind the scenes so users don’t need to navigate multiple apps.
Robust memory systems so the AI can retain and use context over time.
He notes that all major models (Gemini, Claude, Grok, GPT-4.5) are now roughly comparable in performance, with differences being subtle enough that prompting technique matters more than model choice. Khan Academy now relies on its own evaluation framework rather than casual testing.
Evaluating AI Models
Khan Academy maintains several hundred “tough test cases” that historically tripped up models (e.g., distributive property problems, evaluation of student answers like 0.33 vs. 1/3).
When they started, models failed on about 70% of these hard cases; now the failure rate is below 10%.
They use a combination of machine labeling (AI reviewing AI interactions for errors) and human labeling to measure accuracy.
They also label for conversation quality: what percentage of interactions are healthy and engaged versus students disengaged and repeatedly saying “I don’t know.”
In a sample of about 2,000 conversations done roughly six months ago, they found the 2–3% error rate and are working to also measure engagement quality, which they acknowledge likely mirrors traditional classroom dynamics where a significant portion of students are checked out.
Gamification and Student Engagement
Khan Academy has experimented with gamification over the years but has had to align to educational standards, some of which are not inherently fun for students.
He sees promise in quest-based learning, scavenger hunts, and escape-room-style formats, where students will tackle hard cognitive problems if embedded in a broader game.
He personally created a scavenger hunt for his 10-year-old’s birthday using a reasoning model, and believes that within a few years, AI will be able to generate such educational games on its own.
Global Impact and Future Prospects
Khan acknowledges that highly self-moted students anywhere in the world can already benefit from existing tools. He cites a young woman from Afghanistan who couldn’t attend school, used Khan Academy and Schoolhouse.world (a sister nonprofit for volunteer tutoring) to prove her knowledge, and was admitted to MIT.
However, he is skeptical of the techno-optimist view that simply giving students in underserved areas access to ChatGPT or tablets will be transformative. Most students don’t know what they don’t know and can’t structure their own learning journey without guidance.
He believes structured content, like Khan Academy’s, remains essential, but that within a year or two, AI will be good enough that a student with a $30–50 tablet and Starlink access, even sharing the device with five or six other students, could make significant progress.
A major near-term goal: within about two years, Khan Academy hopes to offer high school credits and diplomas, and eventually college credits, which would make the platform dramatically more compelling for students in emerging markets seeking internationally recognized credentials.
The main bottleneck is funding for automated transcription to support multilingual delivery.
The Broader AI and Education Market
Khan sees a lot of noise in the space, with many startups building thin prompting layers on top of ChatGPT and struggling to achieve lasting traction.
He believes Khan Academy’s advantages are: (1) the ability to take a longer view as a nonprofit, with funders willing to support multi-year projects in assessment and writing, and (2) trust built over nearly 20 years around pedagogy, efficacy, and not undermining teachers.
He notes that some investor incentives push startups to seek product-market fit within 9 months or disappear, which is at odds with the slow, trust-building work required in education.
Areas he finds interesting but underdeveloped:
Standardized, AI-driven assessment and credentialing to replace the broken, expensive, and non-standardized interview and hiring process. He envisions a world where passing a final round at one company generates a verifiable certificate that other companies can build on, rather than making candidates start from scratch.
Corporate training, such as interactive AI simulations for workplace scenarios (e.g., “the following situation just played out—what do you do next?”) to replace boring compliance training.
Skills for the Future Workforce
Khan believes foundational skills (reading, writing, math, history, civics) remain essential as a baseline.
The skill that will increasingly matter is what economists call “entrepreneurship as a factor of production”: the ability to take existing resources and combine them in new ways to create value.
In practice, this means people who hear about a new tool or capability and immediately think about how to combine it with what they already know, prototype something quickly (even if it gets them only 80% of the way), and then use their skills to refine it.
He looks for this trait in hiring at Khan Academy and believes organizations that lack this internal entrepreneurship will suffer from higher costs and slower innovation.
Quickfire Round
Changed his mind on: He has shifted further toward believing AI on its own will not solve education’s problems; the key is how AI empowers teachers to hold students accountable and engage them in productive activities.
Favorite way he uses AI: When preparing for videos, he uses AI to ask and answer the “dumb questions” that learners might be afraid to ask, dramatically accelerating his research process. He also uses AI to generate illustrative images for videos and to draft speeches from recorded verbal dumps.
Wish list / experimenting with: “Vibe coding”—using AI platforms to generate and host entire apps from prompts. He is encouraging his 16-year-old son, who is an avid programmer, to embrace this, though his son insists on learning to code properly first.
Biggest surprise in building Khanmigo: The sheer amount of work required to make AI work well at scale in a specific use case, far beyond the initial excitement of the raw model capabilities. On the positive side, both his team and the education community have moved faster than he expected: the company now sees itself as “AI-first,” and the broader education world has embraced AI with less resistance than he anticipated.
The early days after the GPT-4 demo: Khan was so excited that he personally demoed the model to colleagues one by one, many of whom thought they were being pranked. The company was initially split: half saw it as transformative, the other half worried about errors, hallucinations, bias, and cheating. Khan’s response was that both sides were right, but the risks had to be turned into features rather than avoided.
