What Makes an AI Tutor Actually Work: A Field Guide

Khan Academy has one of the most advanced AI tutors on the planet, and by its own admission only about 15% of eligible students ever actually use it. That single number is the most important lesson in AI tutoring right now, and it has almost nothing to do with how smart the model is. Here is what Khan Academy's engineers found when they rebuilt the tutor, what the learning science says about why the fix works, and how to apply both to an AI tutor for workplace training.

The founder of that tutor, Sal Khan, said in July 2026 that the first version, launched three years earlier, "did not change student learning as much as many of us hoped it would." His Chief Learning Officer put it more bluntly: too many students who had it available did not even try it. Then the team rebuilt it, measured everything, and published what moved the needle. The result reads like a design manual for anyone who wants AI to teach, not just answer.

Why do AI tutors get ignored?

AI tutors get ignored when they sit next to the content instead of inside it. A tutor that lives in a chat box off to the side asks the learner to do two hard things at once: notice they are stuck, and then decide to go ask for help. Most people do neither. They keep clicking, or they leave.

The numbers make the problem concrete. Khan Academy's first Khanmigo recorded over 108 million interactions since its 2023 launch, an enormous total that still hid the real issue: engagement was concentrated in a small, motivated minority, while roughly 85% of eligible students never engaged.

Availability is not adoption. A tutor nobody opens teaches nobody, no matter how good its answers are. This is the trap almost every AI learning tool falls into first.

What Khan Academy changed to make its AI tutor work

Khan Academy made its AI tutor work by weaving it into the content, making it proactive, and feeding it each learner's performance data. As the team concluded, the AI "could not just sit next to the content. It had to be woven into it."

In practice, that meant three shifts:

• From reactive to proactive. Instead of waiting to be asked, the tutor now prompts learners to explain their reasoning while they work a problem.
• From generic to personalized. It receives each learner's recent performance and skill progress, so it can target the exact gap instead of starting from zero.
• From bolt-on to built-in. It responds to the specific question in front of the learner, not a blank prompt in a separate window.

The payoff showed up in the data. Across more than 15 million tutoring conversations over six months, giving the tutor a learner's recent problem history improved their next-problem correctness by 3.4%, and surfacing unmastered prerequisite skills added another 2.7%, for a combined 6.1% gain. Structured learning history lifted engagement by about 5%. Even shaving response time by 0.3 to 3 seconds measurably kept learners focused.

Just as telling was what did not work: adding extra example problems or "related content" links produced no measurable improvement. More material is not more learning. Better interaction is.

The learning science: why interaction beats passive video

Interaction beats passive video because the act of retrieving, explaining, and being corrected is what builds durable memory, and decades of research back it up. Watching content is the weakest form of learning there is.

Cognitive scientist Michelene Chi's ICAP framework ranks four levels of engagement: passive, active, constructive, and interactive. Learning rises at every step up that ladder.

• Passive: watching a video or reading a slide.
• Active: clicking, highlighting, or answering a simple question.
• Constructive: explaining your reasoning or generating an answer in your own words.
• Interactive: going back and forth with a tutor that responds to what you actually said.

Khan Academy's own tagging of its tutor conversations as "passive, active, or constructive" comes straight from this research. The goal of a good tutor is to keep nudging learners up the ladder.

The effect is large and well replicated. A landmark meta-analysis of 225 studies found that active-learning methods raised exam scores by 0.47 standard deviations and that students in traditional lecture courses were 1.5 times more likely to fail than those in active ones. Passive delivery is where learning quietly goes to die. That is true for a lecture hall, and it is just as true for an onboarding video playing in a background tab.

How do AI tutors apply to employee training and onboarding?

The same design rules apply directly to employee training: an AI tutor works when it is embedded in the learning moment, adapts to the person, and asks them to think. It fails when it is a chatbot bolted onto your LMS.

Corporate L&D is quietly repeating the exact mistake Khan Academy just corrected. A help chatbot pinned to the corner of a course portal has the same 15% problem: the motivated few use it, everyone else scrolls past. Compliance modules autoplay while new hires check email in another tab. The content is available, and almost nobody engages with it.

The fix is not a smarter side chat. It is building the interaction into the experience itself. That is the premise behind interactive training videos: a video that pauses to ask the learner a question, an AI tutor that responds in real time to their answer, and content that adapts to what each person already knows. Engagement stops being optional because the learning moment and the interaction become the same moment.

The evidence base is catching up to the idea. J-PAL, the poverty-action lab at MIT, is running a randomized controlled trial with Khan Academy that tests exactly this. It compares a tutor students have to summon themselves against one that pops up proactively after a wrong answer, alongside a human coach guiding regular practice. The design encodes the hypothesis every builder should now assume: proactive and embedded beats opt-in.

How to build an AI tutor people actually use

To build an AI tutor people actually use, embed it in the moment of learning, make it proactive, personalize it with performance data, prompt reasoning instead of handing over answers, and measure engagement rather than availability. Here is the sequence:

1. Embed it in the content. Put the tutor inside the video, lesson, or task, not in a separate window. The learner should never have to go find help.
2. Make the first move. Have it prompt at the natural friction point: after a wrong answer, at the start of a tricky section, or when someone stalls.
3. Feed it context. Give it the learner's recent performance and known gaps so it targets the right skill, not a generic explanation.
4. Ask, do not tell. Prompt learners to explain their thinking. The point is to push them up the ICAP ladder from passive to constructive, not to spoon-feed answers.
5. Keep it fast. Latency breaks the illusion of a conversation. Near-instant responses keep people in flow.
6. Measure the right thing. Track what share of learners actually engage and whether they do better on the next task, not how many seats have access.

None of these require a bigger model. They require putting the interaction where the learning happens.

FAQ

Do AI tutors actually improve learning outcomes?

They can, but only under the right conditions. Khan Academy's rebuilt tutor improved next-problem correctness by about 6% once it was embedded and personalized, while its first bolt-on version barely moved learning at all. The design, not the size of the model, decides the outcome.

What is the difference between an AI tutor and a chatbot?

A chatbot answers questions when someone asks. An AI tutor is proactive: it watches how you are doing, prompts you to reason through problems, and adapts to your specific gaps. The difference is roughly the difference between a search box and a teacher.

Why do most AI tutors fail to get used?

Most fail because they sit next to the learning instead of inside it, so people have to notice they are stuck and then choose to ask for help. Only a motivated minority ever does. Tutors that are embedded and proactive close that gap.

The headline from Khan Academy is not that AI tutoring failed. It is that passive delivery fails and interaction works, whether the teacher is a person or a model. The organizations that win the next round of training will not be the ones with the smartest chatbot in the corner. They will be the ones who turn their content from something people watch into something people do.