When Sal Khan first approached OpenAI about building an AI tutor in 2022, the pitch was almost comically simple: take the world's most powerful language model and make it explain algebra to teenagers. What followed, according to a new book excerpt from Josh Tyrangiel, was a two-year experiment that would test every assumption Silicon Valley holds about how technology transforms institutions.
The collaboration produced Khanmigo, an AI assistant now used by millions of students. But the excerpt, published Saturday, suggests the real story isn't about the chatbot's capabilities—it's about the institutional friction that nearly derailed the project at every turn.
The Socratic problem
Khanmigo's designers faced an immediate philosophical challenge: AI tutors are most useful when they provide direct answers, but educational research suggests students learn best when they struggle toward solutions themselves. OpenAI's engineers reportedly spent months fine-tuning the model to ask guiding questions rather than simply solving problems—a constraint that ran counter to everything large language models are optimized to do.
The tension extended beyond pedagogy. Teachers unions expressed concern about job displacement. School administrators worried about liability when students inevitably tried to coax inappropriate responses from the system. Privacy advocates questioned whether children's learning data should flow through servers controlled by a company valued at over $150 billion.
The trust deficit
Perhaps most revealing is how the excerpt describes OpenAI's internal debates about educational deployment. Some executives reportedly pushed for faster rollout, arguing that any AI tutor—even an imperfect one—would outperform the status quo of overcrowded classrooms and undertrained substitute teachers. Others cautioned that a single viral screenshot of Khanmigo giving wrong or harmful advice could set back AI adoption in schools by a decade.
Khan Academy ultimately adopted an approach that satisfied neither camp: limited pilots with intensive human oversight, followed by gradual expansion. The strategy preserved the nonprofit's reputation but frustrated OpenAI's ambitions for rapid scale. By the time Khanmigo reached broad availability, competitors including Google and Anthropic had launched their own educational tools.
The early results
Preliminary data from school districts using Khanmigo shows modest but measurable improvements in student engagement and test scores, particularly among students who previously had limited access to one-on-one tutoring. Critics note that the studies were largely funded by Khan Academy itself and lack the rigor of randomized controlled trials.
What the numbers don't capture is the cultural shift already underway. Teachers report that students now arrive in class having already consulted AI for homework help—whether through Khanmigo or its less educationally-minded competitors. The question is no longer whether AI will reshape learning, but whether purpose-built tools can compete with the general-purpose chatbots students already carry in their pockets.
Our take
The Khanmigo story is less a triumph than a warning. OpenAI and Khan Academy did almost everything right—they prioritized safety, consulted educators, moved deliberately—and still ended up in a market where ChatGPT and Claude are the default study buddies for most students. The lesson for every industry watching AI's advance is sobering: building the responsible product may matter less than building the product people actually use. Education technology has always been littered with well-intentioned failures that lost to whatever was free and convenient. Khanmigo's fate will reveal whether this time is different, or whether the AI tutor joins the interactive whiteboard in the graveyard of pedagogical promise.
