Khan Academy is overhauling its AI-powered tutor, Khanmigo, after admitting the original version did not boost student learning as much as expected. By weaving the AI directly into practice problems rather than leaving it as a separate sidebar, the educational nonprofit hopes to encourage active student engagement. This shift shows a broader challenge in educational technology: moving AI from a tool that simply gives answers to one that drives actual understanding.
What Happened
Sal Khan recently clarified that the first iteration of Khanmigo fell short of expectations, as we previously reported. According to a post on the Khan Academy Blog, many students who had access to the initial chatbot did not use it. To address this, Khan Academy is shifting from "cognitive offloading," where AI simply provides answers, to "cognitive onloading," which forces students to actively work through problems. The updated tool can now see the exact math problem a student is solving, flag mistakes in real-time, and ask the student to explain their reasoning. This update comes as internal product testing from October 2025 to April 2026 revealed a six-percentage-point increase in productive learning interactions.
The Bigger Picture
The struggle to make AI tutors effective is documented in recent academic research. An independent, peer-reviewed study published by the Hermilio Valdizan National University evaluated Khanmigo's impact on secondary school students over eight weeks. The researchers found an increase in mathematical competencies and student motivation, suggesting the tool is effective when students actually engage with it.
However, keeping students engaged without encouraging academic laziness is a difficult balance. A study published in Frontiers in Psychology warns that AI tutoring can lead to superficial thinking if students rely on it to skip the hard work of learning. This study found that while AI can help students solve similar problems shortly after a lesson, this knowledge transfer is modest and drops off significantly within a week. Research in BMC Medical Education also shows that AI learning tools are more effective when guided by a human teacher, showing a Hedges' g = 0.36 advantage compared to completely self-directed use.
To prevent AI from simply giving away answers, computer scientists are designing specific guardrails. A study presented at the ACM Conference on Innovation and Technology in Computer Science Education outlines how Socratic AI architectures use misconception detection to prevent large language models from generating direct answers, forcing a productive struggle that helps students correct their own mental models.
What This Means for Families
For parents and educators, this redesign is a reminder that AI tools are not magic solutions. An AI assistant that sits quietly in the corner of a screen is easily ignored. An AI that acts as a quick shortcut to homework answers can harm long-term retention. Khan Academy’s shift to integrate Khanmigo directly into lessons shows that digital tools require deliberate design to keep kids thinking. Technology is most powerful when it acts as an assistant to human instruction, rather than a replacement.
What You Can Do
When choosing learning apps, opt for programs that ask guiding questions rather than those that instantly solve equations or write essays.
Do not leave children entirely unsupervised with AI tutors. Ask them to explain what the AI is teaching them to reinforce learning transfer.
Teach children that feeling stuck is a normal part of learning. Explain that using AI to bypass difficult problems prevents their brains from building new skills.