When Khan Academy launched Khanmigo in 2023, it put one of AI’s biggest promises for education to the test: a personal tutor for every student.
Khan Academy already had the content, the brand, relationships with schools, and a learning model people trusted. It had early access to GPT-4, built the model around its own materials, and tried to stop it from simply handing students answers.
Sal Khan described the opportunity in his TED talk as the biggest positive transformation education had ever seen. A book followed, and Microsoft later partnered with Khan Academy to make Khanmigo for Teachers available across 49 countries.
That makes the recent reporting on Khanmigo more useful than the usual story about an AI product failing to live up to hype. In April, Khan told Chalkbeat that the tool had been a “non-event” for a lot of students because they did not use it much. A two-year study in 18 Tennessee middle schools found that almost every student tried Khanmigo at least once, but the median student used it in only 17% of the exercise sessions where they had made a mistake. Khan Academy’s chief learning officer, Kristen DiCerbo, later said the company was seeing more passive prompts such as “IDK IDK” than it wanted.
The obvious response is to put the tutor in front of students more often, and Khan Academy has moved in that direction. In its redesigned classroom experience, Khanmigo activates shortly after an exercise page loads rather than waiting for a student to open it.
But the Khanmigo story still points to a deeper assumption in the AI tutor model. It assumes that the learner has already noticed a gap in their knowledge, knows enough to describe that gap, and has a reason to work through the answer. That is often true once somebody has experience. A developer can ask a useful question about an error because they know what the system was supposed to do and can see that it has not done it. Someone learning a new technical field from the beginning may not have that language yet. They might not know that the problem is memory rather than prompting, or an API key rather than the model itself. Before they have tried to make something, they may not know what questions are available to ask.
This is why, at NextWork, we start with building projects. A learner may struggle to say what they need to learn about AI, but once they try to build with it, they encounter problems they can describe: it ignored their document, forgot an instruction, or gave an answer they couldn’t trust. Those problems give them somewhere to begin asking for help.
I do not think the lesson from Khanmigo is that AI tutors failed. I think it is that help-seeking is part of the skill. If we want the tutor to matter, we have to design the moment before the question - not only the quality of the answer that comes after it.
Happy learning,
