Khan Academy's Khanmigo Reaches 10 Million Students

Published 2025-03-22 · Education AI

Khan Academy announced on March 20, 2025, that Khanmigo, its AI-powered tutoring assistant, has reached 10 million registered students worldwide since its public launch in March 2023. The milestone is accompanied by the release of the platform's most comprehensive efficacy data to date: an internal study of 340,000 active users shows that students who engaged with Khanmigo for at least 30 minutes per week improved their maths proficiency scores by an average of 22% over a six-month period, compared with a 9% improvement for Khan Academy users who did not use the AI tutor.

The figures represent a significant inflection point for AI-assisted education. While smaller studies have shown promising results for intelligent tutoring systems, Khanmigo is the first AI tutor to demonstrate measurable learning gains at this scale, across this many languages, and in this diverse a range of educational contexts.

How Khanmigo Works

Khanmigo is built on a customised version of OpenAI's GPT-4 large language model, fine-tuned on Khan Academy's extensive library of educational content and constrained by a pedagogical framework designed to guide rather than give answers. The system operates as a Socratic tutor: when a student struggles with a problem, Khanmigo does not provide the solution. Instead, it asks leading questions, offers hints calibrated to the student's apparent level of understanding, and decomposes complex problems into smaller steps.

This approach — often called "productive struggle" in educational research — is deliberate. Sal Khan, Khan Academy's founder and a former hedge fund analyst, has been vocal about the distinction between an AI tutor and an AI answer engine. "If the model just gives students the answer, it has failed at the one thing that matters most in learning," Khan wrote in a blog post accompanying the announcement. "Khanmigo is designed to keep students in that productive zone of difficulty where actual learning happens."

The system integrates directly into Khan Academy's existing course structure, covering mathematics from arithmetic through multivariable calculus, introductory computer science, biology, chemistry, physics, and US history. In each subject, the AI tutor draws on Khan Academy's pre-existing mastery learning framework, which breaks curricula into discrete skills and tracks student progress through repeated practice and formative assessment.

Language Coverage and Localisation

Khanmigo operates in 42 languages, a fourfold increase from the 10 languages available at launch. The expansion has been driven by a combination of machine translation, community localisation, and partnerships with educational ministries in India, Brazil, Mexico, and Indonesia. In India alone, the platform is available in Hindi, Tamil, Telugu, Bengali, Marathi, and Gujarati, with Urdu and Kannada versions in development.

Localisation extends beyond translation. Khan Academy has adapted example problems, cultural references, and pedagogical sequences to match local curricula. In Brazil, for instance, Khanmigo's maths problems reference Brazilian currency and measurement units, and the pacing of the geometry module has been aligned with the Base Nacional Comum Curricular. In Indonesia, the system incorporates Islamic calendar references in word problems about time and date calculations.

The multilingual capability is powered by a translation-augmented generation architecture that first processes the student's input in the original language, reasons in English (where the model's training data is richest), and then generates a response in the student's language. This approach leverages the model's strongest linguistic capabilities while maintaining the feel of a native-language interaction.

Efficacy Data in Detail

The efficacy analysis was conducted by Khan Academy's internal research team, led by Dr. Marisol Gonzalez, a learning scientist who joined the organisation from the Carnegie Mellon University Human-Computer Interaction Institute. The study examined 340,000 students who used Khan Academy consistently between September 2024 and February 2025, comparing outcomes between those who activated Khanmigo (treatment group, n = 186,000) and those who used Khan Academy's standard practice and video tools without AI tutoring (control group, n = 154,000).

Proficiency was measured using Khan Academy's internal mastery scoring system, which assigns skill-level ratings from "not started" through "attempted," "familiar," "proficient," and "mastered" based on performance on adaptive problem sets. Students in the Khanmigo group progressed an average of 1.8 skill levels over the six-month period, compared with 0.9 skill levels in the control group — a 100% relative improvement in skill progression rate.

On standardised external assessments, the results were more modest but still significant. Among the subset of 12,400 US students who took the NWEA MAP Growth maths assessment both before and after the study period, Khanmigo users showed an average improvement of 8.3 RIT points, compared with 4.9 points for non-users. This difference corresponds to approximately 1.2 additional months of academic growth — a meaningful increment, though not the dramatic leap that some AI education proponents have claimed.

Subgroup Analysis

The data revealed several noteworthy subgroup patterns. Students who had previously been performing below grade level showed the largest gains: those in the bottom quartile of baseline proficiency improved by 31% with Khanmigo, compared with 8% for similar students in the control group. This finding is consistent with the broader educational research consensus that targeted, individualised instruction benefits struggling learners most.

Students from low-income backgrounds — defined as those attending schools where more than 60% of students qualify for free or reduced-price lunch — showed gains roughly equivalent to their higher-income peers, a result that Khan attributed to the platform's accessibility. "A student with a smartphone and a free Khan Academy account gets the same quality of AI tutoring as a student whose family can afford a private tutor," he said. "That democratisation is the whole point."

Older students (grades 9-12) showed somewhat smaller gains than younger students (grades 4-8), which Dr. Gonzalez suggested may reflect the greater complexity of advanced mathematical concepts and the reduced willingness of teenagers to engage in sustained interaction with a conversational AI system.

Teacher Integration and Classroom Use

Approximately 40% of Khanmigo's active users access the platform through school-based programmes rather than independent use. For these students, teachers receive a dashboard showing individual and class-level engagement metrics, common areas of confusion, and recommended intervention points. The teacher dashboard has been available since September 2024 and is now used by approximately 85,000 teachers worldwide.

Khan Academy has positioned the teacher dashboard as central to its institutional strategy. Rather than marketing Khanmigo as a replacement for classroom instruction, the organisation emphasises its role as a "teaching assistant that never sleeps" — handling repetitive tasks like drill practice, homework review, and formative assessment so that teachers can focus on higher-value activities like discussion facilitation, project-based learning, and individual mentoring.

This framing has been effective. In a survey of 4,200 US teachers using Khanmigo, 73% reported that the tool reduced their grading workload, 61% said it helped them identify student misconceptions they would have missed, and 44% reported that it changed their instructional approach in some way. However, 28% expressed concern about over-reliance, and 19% said they had observed students attempting to use Khanmigo to generate homework answers rather than learn from the process.

Looking Ahead

Khan Academy plans to expand Khanmigo into essay writing feedback, science lab simulation, and career counselling later in 2025. The organisation is also exploring a partnership with UNESCO's AI tutoring programmes in developing nations, which have demonstrated similarly strong results in resource-constrained settings. The convergence of these efforts suggests that AI tutoring is approaching the point where it can meaningfully supplement — though not yet replace — traditional instruction at global scale.

The challenge ahead is equitable access. While Khan Academy is free, Khanmigo's most advanced features currently require a subscription of $4 per month or a school-wide licence. Khan has committed to maintaining a robust free tier but has acknowledged that the computational costs of running a large language model at scale make universal free access economically difficult without philanthropic support.

Explore our AI in Education research repository for broader coverage of adaptive learning, intelligent tutoring systems, and education technology policy. Related articles include the UNESCO report on AI tutoring in developing nations and our analysis of WHO's global AI health guidelines.

← Back to Publications