Schools & teaching
AI tutors are getting better. The next question is what students learn.
Khan Academy’s 2026 product tests offer a practical lesson for public AI in American schools.
A student gets unstuck during an AI tutoring session. An answer on the screen is one step. The test comes a moment later: can that student solve the next problem without help?
The need: help that leads to the next independent step
A learner who is stuck needs more than an answer. The next step is to identify the missing concept, work through a smaller example or locate where their reasoning changed direction. Good assistance helps them regain the ability to act.
That is the central design problem for AI tutoring. A system can produce a correct solution while leaving the student unable to solve a similar problem. It can also withhold so much information that the learner becomes frustrated. The useful balance depends on the subject, the learner and the goal of the session.
A public education service needs to make its teaching approach visible. Educators should be able to understand what kind of help is being offered and examine whether the student learns from it.
How Khan Academy tests its tutor
In a May 2026 report, Khan Academy described product tests conducted from October 2025 through April 2026. It evaluated response speed, students’ performance on the next problem without tutoring, and engagement. Some changes to the information supplied to the tutor helped; others did not. The organization also tested faster responses while monitoring mathematical errors and whether the tutor gave answers away. Read Khan Academy’s findings.
Those are developer-reported product experiments, not proof of lasting learning gains across every American classroom. Their value for the American Super Intelligence proposal is the discipline of checking whether assistance leads to independent performance.
There is also a concrete U.S. teaching example. At the July 2025 AFT TEACH conference, educator Tiffany Jefferson demonstrated a subject-focused science chatbot. The session described a progression from teacher use to guided student work and eventual independent use. That earlier example shows why educator preparation belongs alongside the newer technical work. Read the AFT account.
Design the session around the learner’s thinking
In mathematics, a tutoring activity would begin by asking the student to explain their attempted step. If the issue is a misunderstood fraction, the next response would focus on that concept rather than repeat the whole solution. After practice, a new problem would check whether the explanation transferred.
In a science exercise, a learner would predict what will happen in an experiment and explain why. The assistant would ask about evidence and alternative explanations. Actual observation and the teacher’s lesson would remain the basis for resolving the question.
For language learning, a conversation exercise would let someone practice a situation, receive specific feedback and try again. The lesson would distinguish fluency, comprehension and accuracy rather than treat a long conversation as proof of progress.
Make the tutor usable within real teaching
Teachers would need controls over the material, permitted assistance and the information recorded. They would also need a concise way to see where a learner struggled without having to read every exchange. Students should have a straightforward route to human help when a response is confusing or wrong.
Accessibility belongs in the same design. A diagram needs a useful alternative, instructions need to be understandable, and a learner should not be excluded because a preferred interaction method is unavailable. The test is whether the student can participate in the learning activity.
A pilot built around learning
A school pilot should compare supported practice with its existing approach, then assess later recall and transfer to a new problem. It should also measure teacher workload, uneven participation and the time required to correct mistakes. Students and teachers should help explain why the experience did or did not work.
American Super Intelligence has not established a partnership with Khan Academy or the AFT. The proposed contribution is a public foundation that education organizations would evaluate and shape, supported by the skills to use it well.
How Liberty will support a public foundation for tutoring
LibertyGPT will give students access to guided learning, while Liberty OSS will give education teams models they can evaluate and adapt. Subject-specific applications will combine those models with reviewed teaching materials and the controls a school requires.
Multimodal development will expand the kinds of questions a learner can ask about diagrams and documents. Future voice models will add spoken interaction after suitable evaluation. Those capabilities will be introduced according to learning needs and evidence, rather than assumed to make every lesson better.
Education will accompany the tools. Lessons will teach educators to select and review activities and give students practice in checking explanations and working independently. The reason to pursue Liberty is the combination: a public service, adaptable models and shared teaching knowledge that more communities will use and improve.
On the program’s path toward super intelligence, stronger models will need to earn their place in education by helping students understand more and work independently.