The AI Tutor Engagement Gap: Why Access Isn't Enough

AI tutors just beat law professors at their own job. In a blind Stanford Law study of nearly 3,000 answers, faculty preferred the AI response 75% of the time, and flagged it as pedagogically harmful only 3.5% of the time versus 12% for answers written by fellow professors. Yet in the same stretch of 2026, the most famous AI tutor in education got described by its own founder as something most students "just didn't use much."

That contradiction is the AI tutor engagement gap: the tools are now good enough to outperform experts on quality, and learners still barely touch them. The intelligence problem is mostly solved. The engagement problem is wide open. Here is what the latest research shows about why that gap exists, and how L&D teams can close it instead of repeating the same expensive mistake.

Why do AI tutors beat human experts but still fail learners?

AI tutors fail learners because answer quality and learner engagement are two different problems, and the first one is now far easier to solve than the second.

The Stanford result is real and hard to dismiss. Professor Julian Nyarko and colleagues from Yale, NYU, and the University of Chicago ran a blind evaluation in which 16 law professors judged answers to 40 realistic contracts questions without knowing which came from a human and which from an AI. The AI won about three out of four head-to-head matchups.

Nyarko picked law on purpose. "We focused on law precisely because it requires judgment, nuanced reasoning, and the ability to navigate ambiguity, not just factual recall," he said. If AI can hold its own there, most workplace training content is not a stretch.

But Nyarko added the caveat that matters most for anyone buying this technology: "how to implement these tools to most effectively improve student learning is still an open question." A brilliant answer that nobody asks for teaches nobody.

What the Khanmigo reset reveals about AI tutor engagement

Khan Academy learned the engagement lesson the expensive way, and its response is the clearest playbook available right now.

Khanmigo launched three years ago as one of the highest-profile AI tutors in the world. This year founder Sal Khan admitted the first version "did not change student learning as much as many of us hoped it would." His blunter version: "For a lot of students, it was a non-event. They just didn't use it much."

Why? Khan Academy's chief learning officer Kristen DiCerbo named the root cause: "Students aren't great at asking questions well." A tutor that waits for a good question will sit idle, because the learners who most need help are the least able to frame it. Khan compared it to a helper sitting at the back of the class: some students walk up, most never do.

The rebuilt Khanmigo inverts the model in four concrete ways:

• Embedded, not bolted on. It moved out of a separate chat window and into the main practice experience, where the work already happens.
• Context-aware. It can see the exact problem a learner is on and discuss that, instead of starting from a blank prompt.
• Proactive. It initiates help rather than waiting to be asked, which tackles the "students don't ask well" problem head on.
• Reasoning-first. It asks learners to explain how they reached an answer, turning a lookup into a thinking exercise.

Every one of those changes is about engagement, not intelligence. The model did not get smarter. The interaction got designed.

Why access to an AI tutor is not the same as training

Access alone does almost nothing, and there is now hard data on exactly how little.

A 2026 Stanford study from the National Student Support Accelerator, bluntly titled "Access Is Not Enough," tracked what happened when students were simply handed an AI reading tutor. Even with scheduled time to use it, only 61% and 53% of students in the two districts ever logged on. Average use landed at 2 to 5 minutes per week against a platform recommendation of 30 minutes for measurable gains.

Lead researcher Carly Robinson put it in one line: "Access to this AI tutor isn't the same as using it."

The study did find a lever. Adding human support raised engagement by 71 to 80%. But even then, weekly usage crept up only 1 to 4 minutes, and reading scores did not move. The honest takeaway: dropping a capable AI tutor into an environment and hoping people use it is a plan that fails on contact.

For corporate L&D, swap "students" for "new hires" and the picture is familiar. Buying licenses to an AI assistant and announcing it in a kickoff email is the workplace version of the same experiment, and it produces the same two-minutes-a-week result.

How to close the AI tutor engagement gap in workplace training

You close the AI tutor engagement gap by designing the interaction so that learning is the path of least resistance, not an optional detour. The research points to five moves that work.

1. Put the tutor inside the work, not beside it. Engagement collapses when the AI lives in a separate tab. Embed it in the exact task, course, or video the employee is already doing, so using it costs zero extra clicks.

2. Make it proactive. Do not rely on learners to ask good questions. The best designs interrupt at the right moment: a check after a key concept, a nudge when someone stalls, a question before a section people tend to skim.

3. Ask learners to explain, not just receive. Khanmigo's reasoning prompts mirror a well-established finding in learning science: retrieval and self-explanation build durable memory far better than passively reading a correct answer.

4. Adapt to the individual. A tutor that sees recent performance and adjusts difficulty or focus is doing the one thing static content never could. Personalization is the entire reason to use AI here.

5. Keep a human in the loop. The Stanford data is unambiguous that human support multiplies engagement. AI handles scale and patience, while managers and mentors supply accountability and meaning.

This is exactly the bet behind interactive training videos: instead of a passive clip employees half-watch, the video stops, asks, listens, and responds in real time, with an AI tutor built into the experience rather than parked in a sidebar. The engagement is designed in, so it is not left to chance.

What this means for onboarding and L&D leaders

The near-term winners will not be the teams with the smartest model. They will be the teams that design for engagement.

Passive video remains where most training goes to die: people press play, tab away, and click "complete." Adding a chatbot on the side does not fix that, as Khanmigo's first chapter proved at scale. What fixes it is interaction that is embedded, proactive, adaptive, and backed by real humans.

The good news is that the hard part, model quality, is largely handled. The Stanford law results show the raw capability is already there. The work now is design: turning a capable AI into an experience people actually use.

FAQ

Are AI tutors actually better than human instructors?

On answer quality for well-scoped questions, the evidence is trending yes. Stanford's blind study found faculty preferred AI answers 75% of the time and flagged them as harmful far less often than peer answers. The open question is not quality but engagement: whether learners use the tutor consistently enough to benefit.

Why don't employees use the AI training tools we already bought?

Usually because access was the whole plan. Research shows learners given an AI tutor often use it just two to five minutes a week without design and support around it. Engagement rises when the tool is embedded in real work, prompts people proactively, and is reinforced by a manager or mentor.

How is interactive learning different from an AI chatbot?

A chatbot waits for you to ask. Interactive learning drives the exchange: it presents content, then pauses to question, adapt, and respond based on how you answer. That proactive, embedded design is what closes the engagement gap the first wave of AI tutors ran into.

The takeaway

The first generation of AI tutors settled one debate and opened another. The technology can outperform experts on quality, so intelligence is no longer the bottleneck. Engagement is, and engagement is a design problem, not a model problem. The organizations that win at learning will be the ones that stop shipping passive content with an AI bolted on and start building interaction in from the first frame.