Why Active Learning Makes AI Tutors Actually Deliver

Khan Academy gave millions of students a free AI tutor, and roughly 85% of them never really used it. Founder Sal Khan now says the first version "did not change student learning as much as many of us hoped it would," and his team is rebuilding it from the ground up. The fix was not a smarter model. It was one change in behavior: make the tutor pull learners into active learning instead of sitting quietly and waiting to be asked.

That single shift is the difference between an AI tutor that moves the needle and one that becomes expensive wallpaper. And the lesson is not new. Learning scientists have measured it for decades. What is new is that AI can finally deliver active learning to every learner at once, which is exactly where most training programs still fall short.

What went wrong with the first wave of AI tutors

The first wave of AI tutors failed because they were passive: they answered questions but never demanded that the learner do the work. Khanmigo launched three years ago as a chat window sitting next to the lessons. Students could ask it anything. The problem was that most students never did.

Only about 15% of eligible students actively engaged with the tool, according to reporting on the rebuild. Kristen Eignor DiCerbo, Khan Academy's chief learning officer, put it plainly: "Too many students who had it available did not even try it." Across the platform, Khanmigo logged more than 108 million interactions since its 2023 launch, yet the learning gains still disappointed.

The trap is familiar to anyone who has run a corporate training program. Give people an optional resource and most will skip it, because clicking into a tutor requires already knowing you have a gap. The learners who need help most are often the least likely to raise their hand. A tutor that waits to be summoned mostly serves the students who were going to succeed anyway.

Sal Khan's own diagnosis captures the redesign in one line: "The AI could not just sit next to the content. It had to be woven into it."

What is active learning, and why does it beat passive video

Active learning is any approach that makes the learner do something (retrieve, explain, solve, decide) rather than just absorb information, and the evidence that it beats passive delivery is overwhelming. It is the difference between watching someone lift weights and lifting them yourself.

The landmark evidence is a 2014 meta-analysis of 225 studies published in the Proceedings of the National Academy of Sciences. Comparing traditional lectures against active learning in university STEM courses, researchers found that active learning raised exam scores by about 6% (an effect size of 0.47 standard deviations) and cut failure rates from 34% to 22%. Put another way, students in traditional lecture classes were 1.5 times more likely to fail. You can read the meta-analysis on PubMed.

The pattern holds specifically for the watch-versus-do question. Research from Carnegie Mellon on what learning scientists call the "doer effect" found that doing practice while you learn is dramatically more predictive of outcomes than reading or watching more content, and that the relationship is causal, not just correlation. In Koedinger and colleagues' analyses, the benefit of extra practice was more than six times that of extra watching or reading. The takeaway from this doer effect research is blunt: activity causes learning in a way that passive consumption never does.

This is why passive video is where learning goes to die. A polished video can show a learner exactly what good looks like, but recognition is not the same as the ability to perform. The brain files "I have seen this" as familiarity, and familiarity feels like competence right up until the moment you have to do the thing for real.

How AI tutors turn passive content into active learning

AI tutors work when they force the learner to think out loud, adapt to what that learner just did, and give feedback in the moment. That is precisely the redesign Khan Academy landed on after its first version underdelivered.

The rebuilt Khanmigo does three things the original did not:

• It shows up proactively. Instead of waiting inside a chat box, the tutor appears during assignments and offers help while the student is working, so the learners who would never think to ask still get pulled in.
• It asks before it tells. It prompts learners to explain their reasoning on the specific problem in front of them, which turns a passive read into an act of retrieval and self-explanation.
• It adapts to the learner. It uses recent performance and skill progress to decide whether someone needs a fresh challenge or a review of a shaky prerequisite.

Notice that none of this depends on a bigger language model. It depends on design: weaving the AI into the moment of practice so learning becomes something the student does, not something that happens near them. The proactive, adaptive tutor is scheduled to reach participating school districts in the summer of 2026.

The same principle transfers directly to any AI tutor built for the workplace. An interactive training video that stops to ask a warehouse lead how they would handle a spill, listens to the answer, and corrects a misconception on the spot is doing active learning. A talking-head recording of the same policy is not.

What active learning means for corporate training and onboarding

For corporate training, active learning is the fix for the completion-certificate problem: the gap between finishing a course and actually being able to do the job. Most onboarding still runs on passive video and slide decks that new hires click through, nod at, and forget by Friday. The result is people who "completed" training but freeze the first time the situation is real.

Onboarding is where this bites hardest, because time-to-productivity is a number leaders can feel. Every week a new hire spends passively watching content is a week they are not yet contributing, and much of what they watched will not survive contact with the actual job. Active learning compresses that gap by making practice, not exposure, the core of the experience.

The economics finally make sense too. For most of the last decade, active learning at scale meant hiring enough human coaches to sit with every learner, which no L&D budget could support. AI changes that math. An AI tutor can ask every single learner to explain their reasoning, adapt to each one's answers, and give instant feedback, at a cost that does not balloon with headcount. That is the capability Khan Academy is chasing, and it is the same capability that turns onboarding from a video playlist into real practice.

How to build active learning into your training program

You build active learning into training by replacing moments of passive watching with moments that require a response, then adapting based on that response. Here is a practical sequence:

1. Audit for passivity. Go through your existing courses and mark every stretch where the learner does nothing but watch or read for more than a couple of minutes. Those are your dead zones.
2. Insert a question at each dead zone. Follow the doer effect: a short, low-stakes prompt after each chunk of content beats a single quiz at the end. Ask learners to predict, explain, or apply, not just recall.
3. Make the help proactive, not optional. Do not rely on learners to seek support. Surface it during the task, the way the rebuilt Khanmigo does, so the people with the biggest gaps get reached.
4. Adapt to the response. Use what the learner just did to decide what comes next: a harder scenario for someone who nailed it, a prerequisite refresher for someone who stumbled.
5. Measure doing, not clicking. Track whether learners can perform the skill, not whether the video reached 100%. Completion rates hide the exact failure Khan Academy ran into.

Start with your highest-stakes course, the one where getting it wrong on the job actually costs something, and rebuild that first.

FAQ

Is active learning better than watching training videos?

Yes, and the gap is large. A 225-study meta-analysis found active learning raised exam performance and cut failure rates from 34% to 22%, and Carnegie Mellon's doer effect research found that doing practice is many times more effective than watching or reading. Video still has a role for exposure, but on its own it rarely builds skill.

Why did Khan Academy's first AI tutor underperform?

Because it was passive. The original Khanmigo waited for students to ask it questions, and only about 15% of eligible students ever engaged. The rebuilt version reaches learners proactively during practice, prompts them to explain their reasoning, and adapts to their performance.

How can a small L&D team add active learning without more staff?

Use AI to deliver the practice loop that used to require human coaches. An AI tutor can ask every learner to explain their thinking, adapt to each answer, and give instant feedback at a cost that does not rise with headcount, which is what makes active learning finally affordable at scale.

The lesson from Khanmigo's stumble is not that AI tutors do not work. It is that passive tools do not work, no matter how smart the model behind them is. The programs that win are the ones that make every learner do the thinking, get corrected, and try again, which is exactly what interactive learning was built to deliver.