Khan Academy Rolls Out New Gemini-Powered AI Tools for Classrooms

Learn how Khan Academy’s new Gemini-powered Khanmigo tools work, what research says about AI math accuracy, and how schools are governing classroom AI.

Thursday, August 27, 2026

Key Takeaways

  • A two-year randomized trial by the National Bureau of Economic Research found that Khanmigo raised math achievement by 1.3 national percentile ranks. However, these academic gains were similar to traditional Khan Academy practice because students did not actively engage with the AI tutor.
  • The MathemaTikZ benchmark showed that large language models like Gemini 2.0 Flash and GPT-4o achieve only a 73.9% accuracy rate when they generate K-12 mathematical diagrams. The models frequently fail at spatial reasoning and geometric constraints.
  • New state laws require public and community schools to adopt formal AI policies in 2026. These mandates, which include Ohio’s House Bill 96 and Tennessee's Public Chapter 550, will govern how students and staff use artificial intelligence.
  • Khan Academy updated its Khanmigo platform to include Gemini-powered interactive visual aids for math and science. The update also adds a teacher-controlled lesson builder that keeps human educators in charge of assignments.

Khan Academy is launching new artificial intelligence features to assist teachers and students in math and science classrooms. Built in partnership with Google using its Gemini models, these updates bring interactive visual aids and customizable assignment tools to the platform's AI tutor and teaching assistant, Khanmigo. As we previously reported, school tools are shifting toward teacher-focused controls to ensure technology supports, rather than replaces, human instruction.

What Happened

According to the Google Education announcement, six Google.org fellows spent six months working with Khan Academy engineers to integrate Gemini into Khanmigo. The integration adds two main features for the new school year. First, students can use real-time mathematical and scientific diagrams that update as they solve problems. Second, teachers can use an assignment-builder tool to generate, review, and edit practice questions before assigning them. This partnership builds on Google and Khan Academy's broader commitment to design classroom tools.

The Bigger Picture

While adaptive AI tools aim to make classrooms more interactive, independent research suggests these technologies still face technical hurdles. First, generative models struggle to maintain mathematical accuracy. According to the MathemaTikZ benchmark study, top-tier models like Gemini 2.0 Flash and GPT-4o only achieve a 73.9% accuracy rate when creating K-12 mathematical diagrams. The study notes that models struggle with spatial reasoning and geometric constraints. This is supported by an alphaXiv study on math visual aids, which found that current AI workflows require stronger spatial reasoning before they can reliably assist middle schoolers. Research on the Diagram-MMU benchmark also shows that multimodal models have difficulty editing or parsing scientific diagrams accurately.

Second, researchers question how effectively students engage with AI tutors. A two-year randomized trial published by the National Bureau of Economic Research (and shared via EdWorkingPapers) analyzed middle school students using Khanmigo. The study found that while the tool raised math achievement by 1.3 national percentile ranks, these gains were nearly identical to using standard Khan Academy without AI assistance. The primary issue was that the median student only messaged the AI on a third of their practice days, frequently clicking pre-set prompts instead of entering deep math dialogue. However, a smaller study published on Exa showed different results, finding that Khanmigo boosted mathematical competencies and student motivation over an eight-week period.

What This Means for Families

For parents and educators, these developments show that AI tools are supplements rather than replacements for human instruction. While interactive diagrams can make math more visual, their high error rates mean educators must double-check automated figures.

The push for teacher oversight also aligns with changing state laws. As we detailed in our guide on how K-12 leaders are shifting to real governance, school systems are tightening control over classroom technology. According to the AI Tools Bakery policy tracker, states like Ohio (House Bill 96) and Tennessee (Public Chapter 550) now mandate that schools enforce formal AI policies. To comply, many districts are adopting "academic integrity first" frameworks, such as the ethical templates hosted by The Skill Shift, which demand full teacher accountability for all classroom materials. Educators can also consult WeAreTeachers' classroom AI guide to help students understand the boundaries between helpful study assistance and academic dishonesty.

What You Can Do

  • Manually review any visual aids generated by Khanmigo before using them in lesson plans, as generative models frequently make geometric errors.
  • Talk to your students or children about asking detailed questions when using Khanmigo, rather than clicking pre-set automated buttons.
  • Check that any AI-generated materials align with your school district's specific data privacy and academic integrity rules.
Share: