Every AI tutor claims to guide rather than tell. Khanmigo is the only one where that claim survives contact with a determined teenager trying to get their homework done. The system genuinely resists giving answers, redirects with questions, asks the student to explain their reasoning, and holds the line further than any competitor I have tried. That is not a prompt trick, it is a design commitment that costs them satisfaction scores from users who wanted the answer, and they made it anyway.
The learning material explains why, which matters, because a teacher who understands the design will support it rather than route around it. Sitting inside Khan Academy's existing curriculum is the second structural advantage. The tutor knows what the student has covered, where they are in a progression, and what the next thing is. A standalone chatbot has none of that context and has to guess, which is why standalone tutoring assistants drift into generic encouragement.
The integration is the product here more than the model is, and the teacher material is good on how to use that positioning rather than fight it. The teacher tools are the underrated half. Generating a lesson hook, writing differentiated questions, drafting a rubric, summarising where a class is struggling. These are less headline grabbing than the tutor and probably deliver more value per hour, and the training treats them seriously rather than as an afterthought.
The honesty in the material is worth noting too. It states that the tutor makes mistakes, particularly in maths, and tells teachers to expect it. That is more candour than most vendors manage and it earns trust. Where I would push harder is on exactly how often those maths mistakes happen and what they look like.
Arithmetic slips, algebraic steps that are wrong in the middle but arrive somewhere plausible, and confident explanation of an incorrect method are all things I have seen, and a student who does not yet know the material cannot tell. Saying it makes mistakes is accurate. Showing five real examples of the specific failure modes would be considerably more useful, and the material stops one step short. The Socratic approach has a genuine cost that gets less discussion than it deserves.
A student who is stuck, tired and out of time does not need another question, they need to see a worked example and then try again. The system's refusal is right most of the time and wrong some of the time, and knowing which situation you are in is a judgement the tool cannot make. Teachers should know that a frustrated student abandoning the tutor is sometimes the tool working as designed and sometimes it failing, and nothing in the material helps distinguish those. Coverage outside core academic subjects thins out quickly, and the free teacher access being tied to geography leaves a lot of the world paying or going without, which sits oddly with the organisation's mission.
Three point eight. Restraint, curriculum grounding and institutional honesty put this ahead of everything else in the category, and the remaining gap is specificity about how the tutor fails rather than an admission that it does.