Khan Academy's GPT-4 tutor won't just give answers — it'll ask questions

Khan Academy just pulled back the curtain on Khanmigo, a pilot program that stitches GPT-4 into the fabric of its platform as both a virtual tutor and a teacher’s aide. And the early details suggest this isn’t the kind of AI that spoon-feeds answers. Kristen DiCerbo, the nonprofit’s Chief Learning Officer, framed the whole effort around a brutally common classroom problem: every student shows up with wildly different gaps and needs. The promise here is an assistant that can have a real back-and-forth, probing a student’s reasoning with questions like ‘Why did you answer that way?’ rather than just marking things right or wrong.

The pilot, which went into limited testing in 2022 and is now expanding with a public waitlist, is deliberately cautious. The team knows math is a sore spot for large language models—prone to confident-sounding errors—so early participants will be tasked with flagging mistakes. It’s a pragmatic move from an organization that can’t afford to burn trust. Shawn Jansepar, director of engineering, didn’t mince words: before GPT-4, building genuinely tutor-like capabilities was effectively a non-starter. Now, he says, it’s a way to accelerate a roadmap that was years out.

But the teacher-facing tools might be where the rubber meets the road. Ricky Chandarana, a product director, hinted at a feature that would give teachers a real-time snapshot of every student’s progress, a dashboard that could make differentiation less of a logistical nightmare. That’s a direct shot at the administrative grind that eats into actual teaching time. The idea is to let GPT-4 handle the heavy lifting of generating prompts or instructional materials, freeing educators to focus on the human elements of their job.

DiCerbo also touched on an underrated pain point: the dreaded ‘why do I need to learn this?’ question. Her team is betting that GPT-4 can kill that complaint by tailoring its relevance pitch to each kid’s specific interests. It’s a small interaction, but if it works, it chips away at the existential boredom that derails so much self-directed learning. The entire project hinges on whether the model can consistently be a guide on the side, not a crutch.