How Khan Academy brought AI into learning and lesson preparation

The problem.

Students need different kinds of help with the same material. OpenAI’s account describes Khan Academy’s effort to offer individual questions and explanations while also supporting classroom preparation.

What changed.

Khan Academy introduced a limited GPT-4-powered Khanmigo pilot on 14 March 2023. It placed a conversational tutor alongside assigned work and invited teachers and students to help test the experience. The launch account explicitly acknowledged that AI could still make mathematical errors.

As described by OpenAI and Khan Academy and Khan Academy.

What was reported.

On 21 May 2024, Khan Academy announced free U.S. access to its teacher assistant with Microsoft’s support. The release described more than 20 preparation tasks, including creating rubrics, quiz questions, and lesson hooks. Source: Khan Academy

What we take from it.

The valuable design decision is the assistant’s role. A learner trying to understand a topic and a teacher preparing tomorrow’s class have different goals. We would define those separately before designing the interface. In a learning flow, a useful response might be a hint or a question. In a preparation flow, it might be an editable first draft. Both need a clear next step and a way to check the work.

This also applies to internal training and customer education. Imagine a new team member reading an approved operating guide. An assistant could ask them to explain a step, point them toward the relevant section, and help them recognize what they have missed. That is a more specific brief than putting an unrestricted chatbot beside a document. The source material, expected behavior, and boundaries can all be reviewed before the pilot begins.

A good evaluation should keep assistance and independent ability separate. Someone completing an exercise with a helpful answer on screen has not necessarily learned how to do it alone. We would include a later task without assistance, subject-expert review, and feedback from the people doing the work. For preparation tools, we would count editing and checking time as part of the task. Useful output should reduce the total effort required to produce something accurate and ready to use.

How to evaluate a similar idea.

Start with your situation and a question you can test. These are evaluation steps we would discuss before choosing an implementation.

1. 01

   Choose one learning moment

   Start with a recurring point of confusion in training, onboarding, or customer guidance. Define what the person should understand afterward.

2. 02

   Define the kind of help

   Decide when to ask a question, provide a hint, reference a source, or hand the conversation to a person.

3. 03

   Prepare approved material

   Select the information the assistant may use and review access rules before including learner or employee data.

4. 04

   Review with an expert

   Check example conversations for factual errors, misleading explanations, and answers that do too much of the learner’s work.

5. 05

   Measure the lasting benefit

   Assess independent understanding or total preparation effort, including review. Keep satisfaction separate from demonstrated improvement.